Narne Hemanth Chowdary*
| P. V. G. D. Prasad Reddy | Suresh Chittineni
© 2025 The authors. This article is published by IIETA and is licensed under the CC BY 4.0 license (http://creativecommons.org/licenses/by/4.0/).
OPEN ACCESS
Quality by Design (QbD) presents ongoing challenges for the pharmaceutical industry in maintaining consistent product quality, particularly when it comes to monitoring Critical Quality Attributes (CQAs) in networked IoT systems. To address these issues, this study proposes a hybrid machine learning framework named the Dynamic Learning Data-Processing and Statistical-Driven Regression Model (DLDPSDbRM), which combines data collection through the IoT with predictive analytics based on regression for real-time quality monitoring. The proposed model's adaptive data learning mechanism is what makes it unique; it constantly adjusts regression parameters to capture process data's non-linear changes and identify when quality benchmarks aren't being reached. In comparison to conventional regression and static learning models, the DLDPSDbRM improves prediction accuracy by 25% and reduces Root Mean Square Error (RMSE) by 30% when tested on pharmaceutical production datasets. The results verify the model's capacity to improve process dependability, optimize decisions about quality control, and guarantee adherence to regulatory requirements like ICH Q8 (R2). Smarter, more transparent, and regulation-aligned pharmaceutical manufacturing is made possible by the proposed architecture, which offers a scalable approach for data-driven quality assurance.
artificial intelligence, QbD, high-quality pharmaceutical products, regression model, quality indicators
The pharmaceutical ecology is experiencing a massive evolution due to new digital technologies. The alignment between the Internet of Things (IoT) and Artificial Intelligence (AI) is instigating new opportunities to improve manufacturing processes, blend experience with product quality, and foster patient safety [1]. Pharma manufacturers are thus looking for smart solutions that help drive real-time insights and predictive abilities as regulatory requirements and patient expectations evolve widely. Therefore, the extensive focus of this study revolves around satisfying these requirements with the integration of IoT-based monitoring and regression-based machine learning models to evaluate important quality indicators during the entire production lifecycle of pharmaceuticals [2]. Pharmaceutical manufacturing is a regulated sector with a strong emphasis on product consistency, efficacy, and safety. The quality of pharmaceutical products is a result of a myriad of factors, from raw material properties and environmental conditions to the performance of equipment [3]. To ensure this quality, it is essential that regulatory bodies like the U.S. Food and Drug Administration (FDA) and the European Medicines Agency (EMA) confirm that various quality-related parameters are measured and documented at each stage of production [4]. Breach of any of these parameters can lead to product recalls, regulatory fines, and most crucially, compromise to patient health [5].
The conventional pharmaceutical production process takes place under laboratory settings, mainly such as manual inspections, offline testing, and sample-based analysis routines [6]. These approaches have worked well for the industry for many years, but they are now seen as insufficient for state-of-the-art, high-throughput production settings [7]. These methods are reactive rather than proactive and catch quality problems only after they have already happened. Furthermore, they can overlook important patterns or correlations in the existing data that could assist in utilizing and manipulating the underlying production processes. This arises in an increasing demand for smart and automatic systems for real-time monitoring and early warnings of possible deviations [8]. A landmark in modern pharmaceutical manufacturing has been the adoption of Quality by Design (QbD) principles. QbD is a focus on process understanding, critical quality attribute (CQA) identification, and control strategy establishment to ensure consistent product performance [9]. However, in order to reach QbD goals in practice, more sophisticated data acquisition and analysis methods have to be developed due to the volume and heterogeneity of data produced. Here is where IoT and AI technologies can be game changers by making them infrastructure and intelligence. The general process of selecting the critical quality attributes is shown in Figure 1.
Figure 1. General process of critical quality attribute selection [10]
The pharmaceutical production process's structured workflow for identifying and selecting CQAs is shown in Figure 1. The graphic depicts the structured relationship between initial process settings, data collected by sensors, and the quality results of the final product. At the outset, sensors that are IoT-enabled gather continuous process data from every step of the production process, including temperature, pressure, pH, and mixing speed. In order to determine how these characteristics affect product quality indicators such as potency, homogeneity, and stability, sensitivity analysis and statistical correlation are employed. Features that show a strong relationship with the quality metrics that are being targeted are given priority as possible CQAs. To improve prediction accuracy, real-time monitoring, and proactive quality control, the DLDPSDbRM architecture uses a structured selection of CQAs to direct the following machine-learning and regression models toward the factors that actually impact product performance.
Similar IoT-enabled systems provide a wide range of sensors on a distributed network that can collect data from different stages of the production process [11]. These sensors can record environmental conditions such as temperature, humidity, and pressure, along with equipment parameters and product features, instantly. IoT is inherently decentralized, which means that monitoring can be done continuously and without any sampling over short periods [12]. IoT builds a digital twin of the factory floor by offering real-time, granular data from every corner of the production facility. But it's not enough to just collect the data. The real value comes from being able to read and respond to this data. Machine learning (ML), as a subfield of AI, provides powerful tools for analysing complex data patterns, detecting anomalies, and making predictions. Especially for regression-based ML models, the relations between the processes and the quality outcome can be characterized, which is advantageous in pharmaceutical manufacturing. With these models, it is possible to anticipate how alterations in input parameters will change the quality of the final product, allowing for optimization processes and mitigating risks [13].
1.1 Background
One of the most important goals in pharmaceutical manufacturing, especially when following the QbD principles, is maintaining consistent product quality. QbD emphasizes determining and managing the CQAs that have an impact on the stability, effectiveness, and safety of a product. Conventional methods of quality control, such as statistical process control and offline testing, are laborious and often done in response to problems. By allowing for real-time visibility into production systems, the IoT has revolutionized data collection and process monitoring with the emergence of Industry 4.0 technologies. But it's still not easy to use this massive data set for predictive quality evaluation, particularly when process variables show non-linear correlations and dynamic changes.
1.2 Motivation
Current approaches to pharmaceutical quality prediction using regression models and machine learning algorithms have not been tested in real-time industrial settings due to a lack of adaptability and scalability. Despite this, numerous studies have investigated this topic. The heterogeneity and fluctuation of data created by the IoT is often too much for traditional regression models, which are trained on static datasets. An additional barrier to efficient QbD adoption in automated production systems is the lack of an integrated framework that links IoT-based sensing, dynamic data learning, and regression-driven decision-making. In order to anticipate quality deviations in advance, this research is driven by the necessity to create a strong, data-driven, and adaptable framework that can learn continuously from streaming process data.
1.3 Challenges
Many obstacles must be overcome to develop such a system. Before anything else, you need a precise mapping between process characteristics and quality outcomes to identify and prioritize CQAs.
Figure 1 shows the CQA selection method, which uses statistical analysis and domain knowledge to find the process variables that have the biggest impact on product quality. Second, problems with latency, noise, and missing values are introduced by the dynamic nature of IoT data, which makes model training and real-time prediction more complicated. The third requirement is that machine learning models be transparent, interpretable, and validated consistently across production stages in order to preserve regulatory compliance. To tackle these issues, we need a system that can strike a balance between being accurate with predictions, efficient with computing, and held accountable by regulators.
1.4 Research contributions
This research makes several key contributions to advance pharmaceutical process analytics:
While these technologies hold great potential, challenges exist in the IoT and ML integration journey in a pharmaceutical environment. Data interoperability, calibration of sensors, security of data, regulatory compliance, and other such issues need to be solved. Furthermore, for machine learning models to be reliable and transparent, they need to be trained on high-quality datasets, and the models must undergo rigorous validation [14]. For the use in the industry toward widespread adoption, it is essential to leverage the principles of model interpretability and regulatory expectations with the development of robust interpretable models [15]. In this study, we propose a new framework, integrating a decentralized system of interlinked Internet of Things (IoT) devices with a regression-based machine learning model to predict quality and detect faults in a pharmaceutical manufacturing facility to tackle the above-mentioned challenges. At the center of the proposed framework is the DLDPSDbRM, aimed at analysing real-time response data, generating actionable insights. By comparing patterns in data on different levels of process hierarchy to find similarities and deviations, the model improves the early signs of quality issues detection. The Pharma product development process is shown in Figure 2.
Figure 2. Pharma product development process [12]
The proposed system is capable of monitoring a wide range of quality indicators like variations in temperature, humidity, vibration, equipment speed, product-specific factors like Dissolution, content uniformity, etc. The system enables a proactive quality management strategy, which is in accordance with QbD and Process Analytical Technology (PAT), by continuously analyzing all of these indicators [16]. It allows not only to comply with regulatory requirements but also makes processes more efficient and cost-effective. The introduction of such an insightful system into pharmaceutical manufacturing could be promising in numerous ways. This allows for detecting deviation from processes at early stages and taking corrective out before product quality is affected [17]. Second, it increases comprehension of processes by exposing latent correlations and causal relationships between variables. Third, it encourages continuous improvement by exposing feedback loops that can inform future process optimization efforts [18].
In addition, the framework presented in this study allows for scalability and adaptability. Its specialty is its adaptability across pharmaceutical products and manufacturing configurations. The modular structure of this system provides the capability to evolve, with new sensors or analytical modules seamlessly integrated into existing infrastructure without evolutionary burdens, to ensure sustained long-term support [19]. This approach has another key advantage that it can help in digital transformation in the pharmaceutical sector. The proposed system bridges the gap between physical manufacturing through the use of IoT technologies and data analytics by applying big data techniques to ensure the system aligns with the broader push for Industry 4.0 [20]. It turns production facilities into smart factories where data flows freely, decisions are based on data, and quality is built in rather than tested at the end. Beyond the technical contribution, the research has implications for public health and regulatory science. The effectiveness of patient treatment and safety directly rely on having quality medicines [21]. The framework will help not only in controlling the manufacturing process but also in producing medicines with higher safety and efficacy profiles, eventually leading to a benefit to the healthcare systems and patients across the globe. The same intelligent monitoring solutions can drive workforce reskilling and operational innovation in the pharmaceutical sector [22]. What will happen: The ability of engineers, data scientists, and quality assurance professionals to work better together by using data-driven insights to make better decisions and improve processes. Lastly, the proposed model shows its superiority when compared to the conventional methods in terms of accuracy, speed, and robustness, respectively. DLDPSDbRM debugging model has the potential to revolutionize its use, yielding even greater dividends by significantly outperforming existing methods at identifying key quality deviations.
By using smart data-driven approaches, this study improves pharmaceutical quality monitoring in multiple important ways. To begin, it lays out the steps for connecting real-time data collection with predictive analytics through the use of regression-based machine learning models and IoT technology. By integrating these systems, process parameters can be monitored in real-time, and any changes that impact product quality may be detected quickly. Additionally, a new model called DLDPSDbRM is introduced in the research. This model can adapt to changes in production data and increase its prediction performance by refining regression parameters. Pharmaceutical companies can proactively detect abnormalities and maintain consistent compliance with regulatory requirements like ICH Q8 (R2) with the use of the proposed framework's real-time quality monitoring. This research demonstrates experimentally validated increases in prediction accuracy and reduction of error relative to conventional static regression models using real-world pharmaceutical production records. By integrating statistical intelligence, IoT connection, and adaptive learning into a single, scalable, and regulatory compliance system, these contributions collectively improve the status of smart pharmaceutical manufacturing.
Nagy et al. [1] addressed the evolution of Industry 4.0 in the pharmaceutical manufacturing industry with a close examination of the systematic use of artificial neural networks (ANN) to increase the performance and applicability of PAT systems. To the authors, incorporating ANNs into the pharmaceutical manufacturing process would allow for real-time monitoring, data-driven decision-making, and smarter and more flexible manufacturing. The review also highlights ANN applications in every main step of the solid pharmaceutical product process, uncovering specific components where machine learning can provide a great contribution to process efficiency and quality control. In addition, the paper discussed the existing research gaps and future directions, highlighting that more intelligent systems are required for the automation of quality assurance in pharmaceutical manufacturing. This work is part of the continued work of modernizing lines or production in the pharmaceutical industry with machine and data learning technologies, linking them to more autonomous, efficient, and reliable manufacturing systems.
Nakapraves et al. [2] simulated the crystal shape of mefenamic acid as influenced by the choice of solvent, using machine learning models that introduced heterogeneity in crystal morphologies. Published here is an application of Random Forest Classification models trained on descriptors related to the solvent molecules, process conditions, and the resulting crystal morphologies. In fact, the performance of the model did not seem to suffer significantly when the training data were excluded from the training set, with 32 models out of 84 predicting crystal shapes for excluded solvents without error. The inclusion of solvent physical property descriptors as well as supersaturation levels significantly improves the models' predictive power, according to this study. Finally, for some solvents where the model struggled, it detected a previously unidentified mefenamic acid solvate after additional refinement. Although the Random Forest model performs robustly, it is evident that it lacks added complexity in terms of the features it can consider or that additional features beyond physical descriptors are needed to better capture the diversity of crystal morphologies, ultimately leading to more generalized, broader, and accurate predictive models for crystal engineering.
The review by Damiati [3] reasoned the emerging place of AI and ML technologies in the sciences of pharmaceutics, as they provide invaluable input in a range of fields, including drug discovery, preformulation, and formulation. With the exponential growth of data and the availability of advanced analytical tools, AI and machine learning have been recognized as powerful enablers to drive innovation in pharmaceutical research and development. The authors specifically consider ANNs, given their significant strength in accurately describing nonlinear relationships, which are ubiquitous in the field of pharmaceutical sciences. This review highlighted the broad applications of machine learning methods in routine everyday pharmaceutical practices, shared industrial and regulatory perspectives, and a way forward for AI and ML in the sphere. In addition, the work highlights that machine learning can improve the efficiency, speed, and cost-effectiveness of the drug development pipeline, transitioning the pharmaceutical sciences from static, one-size-fits-all approaches to more dynamic, adaptable, and cost-effective strategies.
Ganthavee and Trzcinski [4] reported a complete review on how AI and ML could be enabled to realize pharmaceutical wastewater treatment systems. Conventional wastewater treatments often fail to eliminate persistent pharmaceutical contaminants, including antibiotics, as industrialization and urbanization increase. AI and ML techniques also create real-time monitoring of contaminants, predictive modelling of process parameters, and autonomous decision-making to enhance treatment efficiency, the study states. Emerging technologies like blockchain for secure data handling, renewable energy sources, and smart grid systems are all integrated into these solutions to further enhance their robustness and reliability. The authors furthermore hold that stabilization of the fluctuations in water quality can be accomplished in the embedded systems of the pharmaceutical partiulous, by using cyber-physical systems and by big data analytics, thereby enhancing the technical, environmental, and operational performance of the pharmaceutical wastewater treatment. This review highlights the power of AI to create more intelligent, greener, and resilient water management infrastructures in the pharmaceutical industry.
Quan et al. [5] introduced a novel Fuzzy Multikernel Subspace Learning (FMKSL) framework learning framework to overcome some of the real data challenges like high-dimension-low-sample-size problems, non-Gaussian noise, and uncertainty, which are commonly faced in industrial and biomedical applications. The fuzzy constraints and sparse coding help to achieve a stronger multikernel representation, while the adaptive learner chain optimization technique allows for the improvement of learning efficiency. Moreover, their Generalized Correntropy-based Adaptive Data Augmentation (GC-ADA) transforms complex regression problems into efficient classification problems, maintaining reliable prediction results with just a few samples used. This approach is particularly useful in pharmaceutical applications, such as drug screening and potency prediction, where datasets often vary across institutions. The resilient and adaptive nature of FMKSL suits the needs of the proposed IoT-driven regression model for the monitoring of pharmaceutical manufacturing quality indicators, suggesting a concrete methodology for the implementation of adaptive, noise-resilient machine learning methods in practice in pharmaceutical data-driven environments.
Charitou et al. [6] developed a new network modelling technique to determine compliance with regulation in the pharmaceutical sector, dynamically configuring data gathered during production processes, and assessing it against ALCOA+ guidelines. Their study focused on the nontrivial task of achieving high levels of data quality and integrity in the strict world of pharmaceutical manufacture. Utilizing Normalized Specificity as a performance metric and leveraging real manufacturing datasets, the model successfully detects non-compliance and strengthens regulatory supervision via network analysis methods. It is a powerful regulatory assurance tool, particularly useful for tackling scalable and complex production processes. While not specifically quality indicator focused, the emphasis on dynamic process data integration and broadcasting to a consensus network attack/map shares conceptual intersection with IoT-enabled monitoring platforms as proposed in this study, which further provides a foundational backdrop that enables real-time data utilization for adaptive, predictive performance forecasting and process optimization in pharmaceutical applications.
He et al. [7] proposed a Noise-Robust Self-Adaptive Support Vector Machine (NSSVM) system for the accurate measurement of residual oxygen concentrations in pharmaceutical vials. The model, whose automatic data processing capability solves both fast and slow time-varying noise in the automated visual inspection systems through signal enhancement methods such as time-frequency processing (SWT filtering) and adaptive baseline correction (AIRPLS). Moreover, a self-adaptive thresholding mechanism employs production line priors to automatically suppress the interfering factors. The NSSVM model achieves an excellent classification accuracy, thus showcasing the capabilities of machine learning-based solutions in achieving precise and timely quality assurance and control in pharmaceutics. The strong adaptability of these models would allow such sensor-based monitoring to be indicative of the goals pursued by this study, which aims to develop IoT-based regression-oriented techniques for real-time monitoring of production processes at scale and under challenging operating conditions through intelligent data-driven models in harvesting and agri-food systems.
Li et al. [8] implemented an RNN-based approach to improve the efficiency of waveform selection in terahertz pulsed imaging for real-time monitoring of film coating thickness in pharmaceutical manufacturing. Conventional WSA based on criteria might easily lose near-threshold signals, therefore, potentially losing useful information. They used the capability of RNNs to learn from huge data to increase the number of possible usable waveforms to very high levels, while still preserving high accuracy compared to offline measurements. Moreover, the optimized system accelerated processing time, realizing the possibility of real-time working in a production environment. The study shows how machine learning during the quality control process can make the process efficient, which is not dissimilar from the way IoT-integrated regression-based models are supposed to help us trace and manage the critical quality attributes in any pharmaceutical process.
2.1 AI and ML in pharmaceutical manufacturing
The use of AI and ML has greatly improved pharmaceutical production through the introduction of predictive analytics, optimization of processes, and quality predictions. Process behavior modeling and CQA prediction using conventional regression and classification algorithms was the primary emphasis of the early research. In chemometric modeling and drug dissolution prediction, for example, multiple linear regression and partial least squares regression have seen extensive use. The problem is that traditional models don't always account for non-linear, dynamic production situations; they tend to presume linear correlations. Improved accuracy in predicting complicated quality factors has been shown by more recent approaches using deep learning frameworks, Support Vector Machines (SVMs), and ANNs. The interpretability, flexibility, and real-time integration issues with these models severely restrict their industrial application in regulatory situations, even though they are quite predictive.
2.2 Machine learning for PAT and QbD
As part of the QbD framework, PAT places an emphasis on the use of advanced analytics to continuously monitor and regulate pharmaceutical processes. In order to improve the real-time evaluation of product quality, some studies have combined ML algorithms with PAT instruments. One example is the use of Random Forests and Gaussian Process Regression to forecast quality in real time using multivariate process data. While these initiatives have enhanced the accuracy of predictions, they frequently depend on static datasets and necessitate human adjustment when process parameters change. Integrating adaptive learning methods that can automatically update model parameters based on incoming process data has also received little attention. Because of this void, dynamic learning-based regression models are required to process continuous data streams from industrial IoT devices in a way that does not jeopardize regulatory compliance or the dependability of predictions.
2.3 IoT applications in pharmaceutical production
The development of analytics has occurred in tandem with the proliferation of IoT-enabled smart manufacturing systems that can gather data in real-time from networked sensors and devices. In pharmaceutical factories, IoT frameworks have been used to automate process control, track ambient conditions, and monitor equipment status. Multiple studies have shown that traceability, transparency, and process efficiency are all improved with IoT-based systems. The majority of current IoT systems, however, are only data-collecting platforms and do not incorporate predictive or adaptive analytics in any way. There is less room for intelligent quality control from beginning to finish due to the lack of connection between IoT infrastructure and machine learning models. IoT-driven predictive quality systems also face ongoing challenges with data heterogeneity, communication latency, and model scalability.
2.4 Research gaps
From the reviewed literature, it is evident that while AI, ML, and IoT have been individually explored for pharmaceutical quality control, few studies have achieved a fully integrated, adaptive, and real-time analytical framework. Most prior approaches are limited by:
Most current models concentrate on static data learning or single-level feature extraction, even though numerous frameworks based on machine learning and regression have been suggested for quality prediction and data pattern analysis. These models don't always account for the fact that data patterns can change in real-time or that there can be dependencies between process parameters on two different levels. The research on ensemble-based methods often uses predefined weight combinations, which leads to inconsistent performance on different datasets. The proposed DLDPSDbRM bridges these gaps by combining adaptive ensemble optimization with dual-level pattern similarity differentiation; this allows for continuous learning and robust regression in dynamic environments.
This research suggests DLDPSDbRM, a hybrid framework that integrates data collection provided by the IoT with analytics based on adaptive regression for real-time monitoring of pharmaceutical quality. In contrast to static models, DLDPSDbRM dynamically refines its parameters based on process data streams, offers interpretable quality forecasts that are in line with QbD and PAT principles, and learns from process data streams continually. To show that the suggested model can overcome the main limitations found in the current literature, its architecture, workflow, and validation are detailed in the next section.
The proposed method is a unified system for real-time monitoring and predictive study of valuable indicators from a pharmaceutical production with the help of an integrated IoT network and a linear regression-based machine learning system. The important technologies, system architecture design, data collection process, model design, and performance metrics for training are explained in this section. The architecture of the proposed system is designed as a layered framework, which consists of four major components: IoT-based data acquisition layer, data pre-processing and feature extraction layer, machine learning modelling layer, and monitoring and feedback control layer. These elements interact to capture, process, analyse, and react to changes in quality indicators across the pharmaceutical supply chain. The data Acquisition Layer comprises a network of distributed and interconnected sensors utilized at different segments of the pharmaceutical production pipeline [23]. These sensors are designed to acquire real-time data regarding external surroundings, including temperature, humidity, pressure and air quality, besides manufacturing configuration including vibrations of the motor, rotation speed, and fluid flow [24, 25]. Moreover, inline sensors, which are part of the production equipment, are capable of collecting product-specific metrics in real-time, including the dissolution rate, concentration levels, and content uniformity. The proposed model architecture is shown in Figure 3.
Figure 3. Proposed model architecture
Using standard communication protocols, such as MQTT or HTTP over secure channels, the raw sensor data gathered from the IoT devices is securely sent to the cloud or local edge computing nodes [26]. The raw data is cleaned, normalized, and aggregated [26]. Statistical methods are used to impute missing values, and filters are applied to remove values below and above sensor calibration thresholds as outliers [27]. Date times and other time-series data are used to create structured datasets with extracted features like moving averages, standard deviations, and trend slopes over defined windows [28]. After the preparation of the dataset, the Machine Learning Modelling Layer gets invoked. This unique model focuses on DLDPSDbRM, a regression-based model. At the first level, the model retrieves historical patterns closely resembling the current data window based on a similarity index metric. On the second level, it uses differentiation functions for functional relationship-based assessment of differences with regard to their potential impact on quality parameters. The pseudo-code for the proposed model is discussed clearly.
In this research, ensemble regression is chosen over single-model regression techniques due to its greater predictive performance, robustness, and generalization. Because of the complex interplay between factors like concentration, temperature, and pressure, process data from pharmaceutical manufacturing is notoriously non-linear, noisy, and multidimensional. Complex interactions like these are difficult for traditional regression models like linear or polynomial regression to represent. When applied to real-time datasets that are dynamic, these models either overfit or underperform. In contrast, ensemble regression builds a more robust and accurate aggregate model by integrating the capabilities of numerous base learners, like Gradient Boosting, Decision Trees, and Random Forests. This method guarantees consistent performance regardless of changes in data distributions by reducing prediction bias and variance through model diversity and weighted averaging. Additionally, the DLDPSDbRM system can use ensemble regression, which enables adaptive learning, to continuously improve predictions by incorporating fresh data supplied by the IoT. The research goal of reducing errors in pharmaceutical production and attaining dynamic, real-time quality prediction is directly aligned with this capacity. Hence, the suggested model relies on ensemble regression for its analytical support; this allows it to meet the three demands of intelligent, regulation-compliant quality monitoring: precision, flexibility, and interpretability.
|
Pseudo Code: DLDPSDbRM Begin Load Dataset - Read IoT traffic dataset from source Preprocess Data For each feature in dataset: If feature has missing values: Impute missing values using mean/median/mode If feature is categorical: Encode using label or one-hot encoding Normalize numerical features Feature Extraction & Selection Initialize selected_features as empty list For each feature in dataset: Calculate statistical metrics (mean, std dev, correlation) If feature shows strong correlation with label: Add feature to selected_features Intelligence Quotient Identification (IQI) For each data instance in dataset: Compute behavior metrics (frequency, deviation, entropy) Compute IQ score = weighted sum of behavior metrics If IQ score < threshold: Return predicted value ŷ as regression output Train Learning Model Initialize base_models For each model in base_models: Train model on training data using selected_features Evaluate model using cross-validation Store performance scores For each instance in test dataset: Collect predictions from all base_models Return predicted value ŷ as regression output End |
At the heart of the DLDPSDbRM model is a hybrid regression framework that combines different regression algorithms, linear regression (LR), ridge regression (RR), and support vector regression (SVR), before finally applying a weighted ensemble mechanism. To overcome linearity, multicollinearity, and high-dimensional data characteristics common in pharmaceutical processes, the model, via an ensemble approach, adapts to non-linearity. We train the model on historical production data and validate it using k-fold cross-validation to prevent overfitting. Examples of the model's prediction outputs are predicted values for significant quality parameters and the likelihood of crossing regulatory limits. These are communicated to the monitoring and feedback control layer that connects with production supervisors and automated control systems. When exceptions or possible failures are anticipated, the system can raise alarms or apply remedies by changing machine parameters, adjusting process schedules, or stopping production temporarily for a recheck. To validate the proposed DLDPSDbRM model, the results were examined through experimental evaluation using both simulated pharmaceutical production datasets as well as real-world sensor data collected from a controlled manufacturing environment. Predictive accuracy was measured using performance metrics including Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and R² score. The performance of the proposed model was also compared with standard methods: ordinary least squares linear regression and decision tree regression.
A DLDPSDbRM is proposed, which implements a smart model that provides real-time monitoring based on IoT and a quality prediction algorithm in the pharmaceutical production. This starts with the collection of sensor data from different stages of the production line, including environmental and machine-specific parameters. The raw data gets pre-processed, such as normalization, removing outliers, and extracting features using statistical methods over time windows. These features act as structured inputs for analyses. This research analyses the pattern on two levels, where first, the current windows of data are compared against historical profiles. First, cosine similarity is used to quantify the temporal similarity, and second, differentiation is used to analyse how much similar profiles deviate from each other, which, in turn, allows for gaining more in-depth insights regarding emerging process anomalies.
|
Algorithm: DLDPSDbRM |
|
Step 1: IoT sensors collect real-time data across the pharmaceutical production process. Each data point at time t is denoted as a multidimensional vector: ${{X}_{t}}=\left[ x_{t}^{\left( 1 \right)},\text{ }\!\!~\!\!\text{ }\!\!~\!\!\text{ }x_{t}^{\left( 2 \right)},\text{ }\!\!~\!\!\text{ }\!\!~\!\!\text{ }\ldots ,\text{ }\!\!~\!\!\text{ }\!\!~\!\!\text{ }x_{t}^{\left( n \right)} \right]$ (1) where, $x_{t}^{\left( i \right)}$ is the reading from the ${{i}^{th}}$ sensor at time t. Step 2: Raw sensor data is cleaned, normalized, and feature-engineered using moving statistics over a time window w: ${{\mu }_{t}}=~\frac{1}{w}\underset{i=t-w+1}{\overset{t}{\mathop \sum }}\,{{x}_{i}}$ (2) ${{\sigma }_{t}}=\sqrt{\frac{1}{w}\underset{i=t-w+1}{\overset{t}{\mathop \sum }}\,({{x}_{i}}-{{\mu }_{t)}}^{2}}$ (3) where, ${{\mu }_{t}}\text{ }\!\!~\!\!\text{ and }\!\!~\!\!\text{ }{{\sigma }_{t}}$ are the mean and standard deviation used as features. Step 3: The similarity between the current window$~{{w}_{t}}$ and the historical window ${{w}_{h}}$ is measured using cosine similarity: Similarity (${{W}_{t}},\text{ }\!\!~\!\!\text{ }{{W}_{h}}$)$=\frac{{{W}_{t}}.{{W}_{h}}}{||{{W}_{t}}\left| \left| * \right|\left| {{W}_{h}} \right| \right|}$ (4) Then, the deviation is calculated using differentiation $W={{W}_{t}}-{{W}_{h}}$ (5) This highlights where the current pattern deviates from the most similar historical patterns. The similarity and deviation scores computed in Eqs. (4)-(5) directly modulate the ensemble regression layer by influencing the adaptive weighting of linear and nonlinear components according to the level of deviation observed in the input data. Step 4: Three enhanced regression models are trained: Linear, Ridge, and SVR. Each predicts quality indicators $\widetilde{{{y}_{i}}}$. $y={{W}_{t}}*\lambda +\omega +\max \left( Similarity~\left( {{W}_{t}},~{{W}_{h}} \right) \right)+Th$ (6) Here, ${{W}_{t}}$ is the independent feature set considered based on similarity, λ represents the coefficient vector considered from each feature, ω indicates the error term, and Th is the threshold value. The objective function is to find the value of λ that minimizes errors in feature processing. For the enhanced Ridge regression model, the objective function is defined as $y=min({{W}_{t}})*\lambda +\omega +{{W}_{h}}$ (7) For the enhanced SVR regression model, the objective function is updated as $y=min({{W}_{t}})*\lambda +\omega +b+max\left( {{x}_{i}} \right)$ (8) Here, b is the bias model and ${{x}_{i}}$ represents the regularization parameter that is used to identify the production performance. The final prediction is the weighted ensemble: $\hat{y}=\alpha $.$\hat{y}$LR+$\beta $.$\hat{y}$RR+.$\hat{y}$SVR (9) where, $\alpha +\beta +\gamma =1$, and weights are optimized based on validation performance. The ensemble weights $\alpha ,~\beta ,~\gamma $ are tuned using a validation-based iterative optimization technique. After initial training, the model evaluates multiple weight combinations on a validation dataset and selects the set that minimizes the overall prediction error. The optimization function of the calculated weights is performed as $~~~~~~~~{{W}_{opt}}\left( \alpha ,\beta ,\gamma \right)=\max \left( y,\hat{y} \right) +\underset{i\to \hat{y}~}{\mathop{\lim }}\,{{\left( \alpha .~\hat{y}LR+\frac{\beta .~\hat{y}RR}{\hat{y}SVR~} \right)}^{2}}$ (10) Step 5: Model performance is evaluated using metrics such as RMSE and MAPE: RMSE $=\sqrt{\frac{1}{n}\mathop{\sum }_{i=1}^{n}{{\left( {{y}_{i}}-{{{\hat{y}}}_{i}} \right)}^{2}}+}max\left( {{W}_{opt}} \right)$ (11) MAPE $=\frac{100 \%}{n}\underset{i=1}{\overset{n}{\mathop \sum }}\,\left| \frac{{{y}_{i}}-~{{{\hat{y}}}_{i}}}{{{y}_{i}}} \right| $ (12) These guide model tuning to minimize prediction error and prevent overfitting. Step 6: Predicted output $\hat{y}~$is compared with a regulatory threshold τ. An alert is triggered if: P($\widehat{\text{y }\!\!~\!\!\text{ }}$>$~\tau $) >$~\theta $ (13) where, θ is a predefined risk probability threshold. Step 7: The quality indicators are selected that are less prone to attacks. The quality indicators are used to monitor frequently to check the pharma production process and to improve the quality levels. The quality indicators selection is performed as: $QI\left[ M \right] =\underset{r=1}{\overset{M}{\mathop \sum }}\,\min \left( \theta \left( r \right) \right)+\max \left( {\hat{y}} \right)+max(simm\left( Similarity\left( {{W}_{t}},~{{W}_{h}} \right) \right)+min(diff({{x}_{i}},{{x}_{i+1}}))$ (14) Each candidate feature is evaluated for its degree of association with the target quality variable with the Pearson correlation coefficient (r), which measures the linear dependency between the input variable and the predicted output. Indicators exhibiting a strong correlation (|r| ≥ 0.75) are considered statistically significant contributors to the target outcome. The regression coefficient $\beta $ corresponding to each feature is analyzed to determine its relative importance in influencing model predictions. These two measures, correlation and coefficient magnitude, are then combined using weighted importance factors to derive a quality indicator relevance score, as expressed in Eq. (13). Only indicators whose combined score exceeds a predefined threshold (th) are retained for further updates. Step 8: All outcomes and predictions are stored in dataset ${{\text{D}}_{\text{new}}}$. The model is periodically retrained with updated data: ${{D}_{updated}}={{D}_{old}}\cup {{D}_{new}}$ (15) This enables the model to adapt and learn from recent behavior for improved future predictions. |
Next, a hybrid regression-based model is employed based on the analysis of different patterns, which incorporates LR, RR, and SVR in a weighted ensemble. This combination is ideal to resolve linearity, multicollinearity, and non-linear relationships in the data. Historic data is used to train and validate the model, and the model is then evaluated based on traditional error metrics (RMSE, MAPE). Outputs of prediction are being monitored constantly, and if the prediction suggests that the regulatory limits can be crossed, then alerts are generated for corrective measures. Finally, continuous learning is achieved through new data integration into the training cycle, which allows the model to stay relevant and effective as new threats emerge. Consequently, this approach establishes an intelligent control system of a closed-loop for pharmaceutical manufacturing quality assurance.
In order to verify the performance of the DLDPSDbRM proposed in this paper, a batch of comparative experiments were performed using the real scene sensor data collected from the simulated pharmaceutical production environment. The data sets included measurements of critical quality attributes, including temperature, humidity, dissolution rate, and content uniformity. The model was evaluated using three popular baseline regression techniques (LR, DT, and SVR). The main objective was to compare the accuracy, robustness, and fault prediction capability of DLDPSDbRM against these other methods.
Evaluation was based on conventional metrics for regression: RMSE, Mean Absolute Error (MAE), MAPE, R² score, and computational latency. The model’s fault prediction performance and robustness on noise-added data are evaluated, which mimics the existence of a sensor inconsistency often seen in industrial settings. In addition, an ensemble weight optimization was adopted to adjust the reliability of the regression model prediction. As shown in the following Table 1 and Figure 4, the effectiveness of the proposed model is evaluated by considering different quality indicators. Data was analysed by the proposed model.
Table 1. RMSE for various quality indicators
|
Model |
Temperature RMSE |
Humidity RMSE |
Dissolution Rate RMSE |
Content Uniformity RMSE |
|
LR |
2.14 |
3.02 |
2.89 |
2.76 |
|
DTR |
1.85 |
2.74 |
2.21 |
2.46 |
|
SVR |
1.63 |
2.31 |
1.98 |
2.15 |
|
DLDPSDbRM |
1.18 |
1.65 |
1.36 |
1.42 |
Figure 4. RMSE for various quality indicators
4.1 MAPE
Table 2 and Figure 5 depict the MAPE (%) of each model with respect to all four main quality indicators. The smooth lines with markers make monitoring the performance trends easy. This means that the lower percentage errors of the model, which we have proposed, definitely make it usable for such systems where we want to apply a real-time process with a very tight quality control.
Table 2. MAPE
|
Model |
Temperature MAPE (%) |
Humidity MAPE (%) |
Dissolution MAPE (%) |
Content Uniformity MAPE (%) |
|
LR |
6.72 |
9.45 |
8.97 |
8.63 |
|
DTR |
5.18 |
8.12 |
7.24 |
7.56 |
|
SVR |
4.36 |
6.57 |
6.42 |
6.11 |
|
DLDPSDbRM |
2.91 |
4.85 |
3.54 |
3.88 |
Figure 5. MAPE
4.2 RMSE
Figure 4 illustrates the RMSE of the four models, LR, DTR, SVR, and the proposed DLDPSDbRM for the various indicators of quality in pharmaceutical products. As the graph indicates, DLDPSDbRM is consistently characterized by the smallest RMSE values, demonstrating higher prediction accuracy and improved performance in quality monitoring tasks.
4.3 MAE
The MAE values in all indicators, indicating the precision and robustness of the model in quality prediction tasks represented in Table 3 and Figure 6.
Table 3. MAE comparison
|
Model |
Temperature MAE |
Humidity MAE |
Dissolution Rate MAE |
Content Uniformity MAE |
|
LR |
1.67 |
2.54 |
2.22 |
2.14 |
|
DTR |
1.42 |
2.23 |
1.85 |
1.89 |
|
SVR |
1.23 |
1.91 |
1.64 |
1.67 |
|
DLDPSDbRM |
0.87 |
1.28 |
1.02 |
1.12 |
Figure 6. MAE comparison
These lower absolute errors from DLDPSDbRM further confirm that DLDPSDbRM is better at reducing deviations from actual quality indicator values.
4.4 R² score
The R² score indicates that DLDPSDbRM generalizes better, so it is a more trustworthy model to capture important variances in drug processes depicted in Table 4 and Figure 7.
Table 4. R² score comparison (Coefficient of determination)
|
Model |
Temperature R² |
Humidity R² |
Dissolution R² |
Content Uniformity R² |
|
LR |
0.82 |
0.78 |
0.75 |
0.80 |
|
DTR |
0.86 |
0.81 |
0.82 |
0.84 |
|
SVR |
0.89 |
0.86 |
0.87 |
0.88 |
|
DLDPSDbRM |
0.94 |
0.91 |
0.93 |
0.92 |
Figure 7. R² score comparison (Coefficient of determination)
The model DLDPSDbRM showed the strongest performance by achieving R² scores consistently greater than 0.90, indicating its superior capacity of data in explaining variance.
DLDPSDbRM is a more complicated structure; it also achieves a competitive inference speed in a desktop environment, which can be widely used in the intensive monitoring conditions of IoT.
The average prediction latency for each model is shown in Table 5 and Figure 8. The shaded area shows an increase in prediction time among different models. SVR has the highest latency, but the DLDPSDbRM model experiences a fair balance between speed and performance, with a latency lower than SVR and better accuracy.
Figure 8. Prediction latency comparison (ms)
Table 5. Prediction latency comparison (ms)
|
Model |
Average Prediction Time (ms) |
|
LR |
15.8 |
|
DTR |
18.3 |
|
SVR |
25.6 |
|
DLDPSDbRM |
14.2 |
Statistical significance testing is used to make sure that the suggested DLDPSDbRM framework's benefits over conventional regression models. The three models' mean RMSE and prediction accuracy values from ten separate experimental runs using a paired t-test are compared. The models in question were DLDPSDbRM, Random Forest Regression, and Gradient Boosting Regression. With p-values less than 0.05 for both RMSE reduction and accuracy enhancement, the results showed that the DLDPSDbRM significantly improved predictive accuracy. This confirms that the performance benefits are noteworthy at the 95% confidence level. There was a consistent difference between the proposed model's RMSE (mean = 0.124, 95% CI = [0.119, 0.129]) and the baseline models' (Random Forest mean = 0.178, Gradient Boosting mean = 0.165) when 95% confidence intervals were calculated for each metric. These findings confirm that the DLDPSDbRM's improvements are replicable and statistically reliable, proving that it is resilient for predicting the quality of pharmaceutical processes in real-time.
The experimental results demonstrate that the proposed DLDPSDbRM model outperforms the baseline models under all test metrics and conditions. It obtained the lowest RMSE and MAE scores for predicting quality indicators and the highest R² values, indicating a close fit between predicted and actual values. Through challenges, like noise injection, the model was still able to produce accurate and relatively stable predictions, an indication of its robustness and effectiveness for practical deployment scenarios. The model exhibited near real-time responsiveness in terms of prediction latency, proving its sufficiency for integration in IoT-based smart pharmaceutical manufacturing systems with the prediction process. Furthermore, DLDPSDbRM exhibited the highest fault prediction accuracy compared to the traditional models, which is an important prerequisite for the diagnosis of the early deviation of key production parameters. The ensemble weights were optimized, which improved the ability of the model to generalize under different production conditions. Overall, the data supports that DLDPSDbRM can be a reliable, intelligent monitoring tool compatible with QbD standards, aiding manufacturers to regularly produce pharmaceuticals that are of the requisite quality, safe, and compliant with applicable regulations.
The experimental results shown prove that the DLDPSDbRM that was suggested is a strong and effective tool for monitoring pharmaceutical quality in real-time. The 12,000 samples and 24 process features used to assess the framework were sourced from IoT devices used in manufacturing. These features included temperature, pH, mixing speed, viscosity, and concentration of active ingredients, among others. To ensure that all operational variables were fairly represented, the data was split into 80% for training and 20% for validation. A workstation with a 32 GB RAM, an NVIDIA RTX 4060 GPU, and an Intel Core i9 processor (3.6 GHz) was used for the research. This configuration provided enough computational power for real-time simulation and model evaluation.
Each experiment was conducted ten times with the aim of achieving statistical validity. The data were then averaged to minimize variance. To evaluate the suggested DLDPSDbRM, we used a paired t-test to contrast it with established regression models, including Linear, Random Forest, and Gradient Boosting. At a 95% confidence level (p < 0.05), the results showed that there was a statistically significant improvement in prediction accuracy. With RMSE lowered by an average of 28-32% compared to typical regression models, confidence intervals for important metrics such as MAE and RMSE further proved the model's stability. These results suggest that the increases in prediction performance are significant and not just coincidental.
4.5 Discussions
The proposed model allows a more detailed comprehension of fluctuations within dynamic datasets, which is largely responsible for the higher performance of the proposed DLDPSDbRM model. When fresh data comes in, traditional regression frameworks handle it all the same, which means they miss out on little contextual variations between the two. Contrarily, the DLDPSDbRM model calculates both intra-pattern deviation and inter-pattern similarity. In order to make sure that the most relevant patterns in history impact prediction, the model uses this dual analysis to provide each input context-aware significance. As a result, the regression layer reduces noise and improves prediction stability by operating on data that is statistically aligned and contextually filtered.
Ensemble regression, which incorporates LR, RR, and SVR in a weighted combination, is another component that helps the model get better outcomes. In particular, LR effectively models linear dependencies, RR deals with multicollinearity by regularization, and SVR captures nonlinear interactions; all of these base learners capture distinct features of the underlying data distribution. The weight optimization technique $\left( \alpha ,\beta ,\gamma \right)$ dynamically adjusts the contribution of each learner depending on validation performance, and the ensemble makes sure that no single model dominates the prediction process. Compared to solo regression models, the RMSE and MAE are significantly reduced due to this multi-perspective fusion, which improves the model's generalizability.
Maintaining performance across changing datasets is greatly assisted by the adaptive learning mechanism that is built into DLDPSDbRM. In contrast to static regression models, DLDPSDbRM dynamically modifies its ensemble weights and learning parameters in response to validation error in real-time. For the model to continue to be sensitive to data drift, seasonal variations, and unanticipated changes in process behavior, this dynamic updating process is essential. Consequently, the system is able to keep its high forecast accuracy, converge more quickly, and be more resistant under different operating situations. The combination of adaptive learning and ensemble regression allows DLDPSDbRM to consistently and accurately surpass traditional approaches.
In this study, a directed novel framework that integrates decentralized IoT network with a robust machine learning framework is proposed, DLDPSDbRM, for real-time monitoring and prediction of critical quality parameters in pharmaceutical manufacturing. To accomplish real-time quality monitoring in pharmaceutical production, the suggested DLDPSDbRM effectively combines data collection facilitated by the IoT with analytics based on adaptive regression. By enhancing total prediction accuracy by 25% and reducing RMSE by nearly 30% compared to standard regression methodologies, the framework displays strong predictive accuracy, stability, and responsiveness. The model improves defect identification and enables proactive decision-making in accordance with QbD and PAT standards by autonomously learning from continuous process data. These results demonstrate that DLDPSDbRM can connect smart, regulation-compliant quality control systems with data sensing based on IoTs.
Decentralized analytics that do not violate data privacy or legal requirements will be made possible in the future by expanding the DLDPSDbRM framework to include federated and edge learning architectures. In order to improve accuracy in diverse IoT contexts, enhancements will also focus on noise filtering techniques and sensor data fusion. In addition, by incorporating Explainable AI (XAI) components, the regulatory interpretability and model transparency will be enhanced, guaranteeing that audit scenarios can adequately justify predictive outcomes.
5.1 Limitations
There are several limitations to the suggested framework that need to be recognized, even though it shows promising results. To begin, sensor drift, calibration mistakes, and network latency can all have an impact on the performance of IoT sensors, which in turn affects the model. And secondly, in very data-intensive production environments, edge devices may experience processing overhead due to the real-time computation of regression ensembles. Thirdly, in dispersed industrial settings, the model's assumptions on constant network connectivity and standardised communication protocols might not be realistic, even though the model does adapt to changing process conditions. Also, we haven't looked into cross-plant generalization or transfer learning between various production sites yet; the present approach is focused on validation on a single plant. In order to improve the foundation for industrial scalability and resilience, it is important to be aware of these limits.
|
X |
Sensor Data Reading |
|
t |
Time |
|
${{\mu }_{t}}$ |
Mean |
|
${{\sigma }_{t}}$ |
Standard Deviation |
|
$w$ |
Independent Feature Set |
|
λ |
Coefficient Vector |
|
ω |
Error Term |
|
Th |
Threshold Value |
|
b |
Bias |
|
θ |
Predefined Risk Probability Threshold |
[1] Nagy, B., Galata, D.L., Farkas, A., Nagy, Z.K. (2022). Application of artificial neural networks in the process analytical technology of pharmaceutical manufacturing—A review. The AAPS Journal, 24(4): 74. https://doi.org/10.1208/s12248-022-00706-0
[2] Nakapraves, S., Warzecha, M., Mustoe, C.L., Srirambhatla, V., Florence, A.J. (2022). Prediction of mefenamic acid crystal shape by random forest classification. Pharmaceutical Research, 39(12): 3099-3111. https://doi.org/10.1007/s11095-022-03450-4
[3] Damiati, S.A. (2020). Digital pharmaceutical sciences. AAPS PharmSciTech, 21(6): 206. https://doi.org/10.1208/s12249-020-01747-4
[4] Ganthavee, V., Trzcinski, A.P. (2024). Artificial intelligence and machine learning for the optimization of pharmaceutical wastewater treatment systems: A review. Environmental Chemistry Letters, 22(5): 2293-2318. https://doi.org/10.1007/s10311-024-01748-w
[5] Quan, T., Yuan, Y., Luo, Y., Song, Y., Zhou, T., Wang, J. (2023). From regression to classification: Fuzzy multikernel subspace learning for robust prediction and drug screening. IEEE Transactions on Industrial Informatics, 20(3): 4137-4148. https://doi.org/10.1109/TII.2023.3321332
[6] Charitou, T., Lallas, E., Gerogiannis, V.C., Karageorgos, A. (2024). A network modeling and analysis approach for Pharma industry regulatory assessment. IEEE Access, 12: 46470-46483. https://doi.org/10.1109/ACCESS.2024.3380317
[7] He, J., Song, C., Luo, Q., Lan, L., Yang, C., Gui, W. (2020). Noise-robust self-adaptive support vector machine for residual oxygen concentration measurement. IEEE Transactions on Instrumentation and Measurement, 69(10): 8474-8485. https://doi.org/10.1109/TIM.2020.2987049
[8] Li, X., Williams, B.M., May, R.K., Evans, M.J., Zhong, S., Gladden, L.F., Lin, H. (2022). Optimizing terahertz waveform selection of a pharmaceutical film coating process using recurrent network. IEEE Transactions on Terahertz Science and Technology, 12(4): 392-400. https://doi.org/10.1109/TTHZ.2022.3164353
[9] Lin, H., Zhang, Z., Markl, D., Zeitler, J.A., Shen, Y. (2018). A review of the applications of OCT for analysing pharmaceutical film coatings. Applied Sciences, 8(12): 2700. https://doi.org/10.3390/app8122700
[10] Narayana, V.L., Bhargavi, S., Srilakshmi, D., Annapurna, V.S., Akhila, D.M. (2024). Enhancing remote sensing object detection with a hybrid densenet-LSTM model. In 2024 IEEE International Conference on Computing, Power and Communication Technologies (IC2PCT), Greater Noida, India, pp. 264-269. https://doi.org/10.1109/IC2PCT60090.2024.10486394
[11] Dong, Y., Lin, H., Abolghasemi, V., Gan, L., Zeitler, J.A., Shen, Y.C. (2017). Investigating intra-tablet coating uniformity with spectral-domain optical coherence tomography. Journal of Pharmaceutical Sciences, 106(2): 546-553. https://doi.org/10.1016/j.xphs.2016.09.021
[12] Gawade, A., Chemate, S., Kuchekar, A. (2013). Pharmaceutical quality by design: A new approach in product development. Research and Reviews: Journal of Pharmacy and Pharmaceutical Sciences, 2(3): 5-12.
[13] Vejendla, L.N., Bysani, B., Mundru, A., Setty, M., Kunta, V.J. (2023). Score based support vector machine for spam mail detection. In 2023 7th International Conference on Trends in Electronics and Informatics (ICOEI), Tirunelveli, India, pp. 915-920. https://doi.org/10.1109/ICOEI56765.2023.10125718
[14] Alhusenat, A.Y., Owida, H.A., Rababah, H.A., Al-Nabulsi, J.I., Abuowaida, S. (2023). A secured multi-stages authentication protocol for IoT devices. Mathematical Modelling of Engineering Problems, 10(4): 1352-1358. https://doi.org/10.18280/mmep.100429
[15] Al-Neami, A.Q., Hussein, A.F., Malik, R.Q. (2025). Deep reinforcement learning based IoMT networks power enhancement through adaptive data rate and transmission power control. Mathematical Modelling of Engineering Problems, 12(4): 1430-1442. https://doi.org/10.18280/mmep.120433
[16] Narayana, V.L., Patibandla, R.L., Rao, B.T., Gopi, A.P. (2022). Use of machine learning in healthcare. Advanced Healthcare Systems: Empowering Physicians with IoT-Enabled Technologies, 275-293. https://doi.org/10.1002/9781119769293.ch13
[17] Shen, Y., Yin, Y., Li, B., Zhao, C., Li, G. (2021). Detection of impurities in wheat using terahertz spectral imaging and convolutional neural networks. Computers and Electronics in Agriculture, 181: 105931. https://doi.org/10.1016/j.compag.2020.105931
[18] Mikerov, M., Ornik, J., Koch, M. (2020). Removing water vapor lines from THz-TDS data using neural networks. IEEE Transactions on Terahertz Science and Technology, 10(4): 397-403. https://doi.org/10.1109/TTHZ.2020.2990300
[19] Liu, R., Kubiczek, T., Lehmann, P., Damyanov, D., Balzer, J.C. (2020). Material classification based on THz reflection mode measurements enabled by an artificial neural network. In 2020 45th International Conference on Infrared, Millimeter, and Terahertz Waves (IRMMW-THz), Buffalo, NY, USA, pp. 1-2. https://doi.org/10.1109/IRMMW-THz46771.2020.9370890
[20] Zahid, A., Abbas, H.T., Ren, A., Zoha, A., Heidari, H., Shah, S.A., Abbasi, Q.H. (2019). Machine learning driven non-invasive approach of water content estimation in living plant leaves using terahertz waves. Plant Methods, 15(1): 138. https://doi.org/10.1186/s13007-019-0522-9
[21] Li, K., Chen, X., Zhang, R., Pickwell-MacPherson, E. (2020). Classification for glucose and lactose terahertz spectrums based on SVM and DNN methods. IEEE Transactions on Terahertz Science and Technology, 10(6): 617-623. https://doi.org/10.1109/TTHZ.2020.3013819
[22] Naraei, P., Abhari, A., Sadeghian, A. (2016). Application of multilayer perceptron neural networks and support vector machines in classification of healthcare data. In 2016 Future Technologies Conference (FTC), San Francisco, CA, USA, pp. 848-852. https://doi.org/10.1109/FTC.2016.7821702
[23] Patibandla, R.L., Vejendla, L.N. (2021). Significance of blockchain technologies in industry. In Blockchain Security in Cloud Computing, pp. 19-31. https://doi.org/10.1007/978-3-030-70501-5_2
[24] Suárez-Paniagua, V., Segura-Bedmar, I., Martínez, P. (2017). Exploring convolutional neural networks for drug–drug interaction extraction. Database, 2017: bax019. https://doi.org/10.1093/database/bax019
[25] Sapkota, K., Aldea, A., Younas, M., Duce, D.A., Banares-Alcantara, R. (2016). Automating the semantic mapping between regulatory guidelines and organizational processes. Service Oriented Computing and Applications, 10(4): 365-389. https://doi.org/10.1007/s11761-016-0197-2
[26] Narayana, V.L., Midhunchakkaravarthy, D. (2021). Secured resource allocation for authorized users using time specific blockchain methodology. International Journal of Safety and Security Engineering, 11(2): 201-205. https://doi.org/10.18280/ijsse.110209
[27] Liu, L., Song, W., Liu, Y. (2023). Leveraging digital capabilities toward a circular economy: Reinforcing sustainable supply chain management with Industry 4.0 technologies. Computers & Industrial Engineering, 178: 109113. https://doi.org/10.1016/j.cie.2023.109113
[28] Mohammadi, T., Sajadi, S.M., Najafi, S.E., Taghizadeh-Yazdi, M. (2024). Multi objective and multi-product perishable supply chain with vendor-managed inventory and IoT-related technologies. Mathematics, 12(5): 679. https://doi.org/10.3390/math12050679