Automated Multimodal Edge Intelligence for Tuberculosis Screening in Low-Resource Settings: A Review and Deployment-Readiness Taxonomy

The global burden of tuberculosis (TB) continues to demand screening approaches that require less infrastructure than facility-based pathways, so that remote and marginalised populations can be reached. Active case-finding with ultraportable radiography and artificial intelligence (AI) shows that screening can be carried into remote settings, yet it also exposes the operational difficulty of establishing such services in the field [1].

Finding missed cases is central to TB control, because individuals outside established diagnostic pathways require faster, more affordable, and more accessible routes to presumptive diagnosis [2]. The World Health Organization (WHO) recognises symptom screening, chest radiography, computer-aided detection (CAD) software, WHO-approved rapid molecular tests, and C-reactive protein as complementary screening and triage tools [3]. The 2025 WHO policy statement on CAD reinforces this framework: six CAD products were found to meet WHO performance standards for TB screening in people aged 15 years and older, although CAD is not yet recommended for children and adolescents younger than 15 [4]. The Stop TB Partnership provides complementary operational guidance on implementing CAD with ultra-portable X-ray systems in screening and triage programmes [5]. An automated screening system should therefore connect screening, triage, and confirmatory testing rather than terminate at a single model prediction.

This is typically done in conjunction with a symptom history, chest X-ray, and confirmatory microbiological investigations. While symptom screening is low cost, it can yield a low detection rate of the disease and a high rate of false positives; AI-assisted chest radiography can therefore improve triage in community screening [6]. Analyses of digital chest X-ray images in hospitalised groups also suggest that AI can aid in triage, but that the benefits rely on integration with confirmation testing [7]. It is crucial to validate externally, choose a high threshold for sensitivity, and consider the population context-in the multi-site study, a TB-detecting radiographic AI achieved a sensitivity of 87% and specificity of 70% at a high threshold; neither the AI nor the radiologists met the WHO threshold for sensitivity in that study population [8].

Apart from radiology, cough-based analysis is an important low-cost channel because cough can be recorded on mobile phones before radiology is available [9]. The most rigorous multi-country evidence to date comes from the CODA TB DREAM Challenge, in which algorithms developed on 2,143 adults from seven high-burden countries showed that cough-only models were modest (area under the receiver operating characteristic curve (AUROC) of 0.69-0.74), whereas cough-plus-clinical models performed markedly better (AUROC 0.78–0.83; best 80% sensitivity and 73.8% specificity), approaching but not yet meeting the WHO target accuracy for a TB screening test [10]. Against this benchmark, a single high-burden-setting study using speech foundation models reported an AUROC increase from 85.2% to 92.1%, with sensitivity of 90.3% and specificity of 73.1% at a representative threshold, although the authors stated that the model was still awaiting broader validation [9].

Multimodal fusion in radiographic settings is supported by early evidence: a single-centre retrospective study reported that combining clinical data with a deep-learning-based chest X-ray detection algorithm yielded an AUROC of 0.924, with specificity of 81.4% at a sensitivity of 90%; given the single-centre design and limited sample, this result should be read as preliminary evidence of feasibility rather than established proof of a generalisable multimodal benefit [11]. Vision-language models (VLMs) can be built with this as well, connecting imaging features and clinical text, but with a computational cost that should be considered in deployment [12].

The issues addressed in the present review, therefore, do not only involve diagnostic accuracy but also deployment. Studies on user acceptance of CAD indicate that user acceptance is influenced by workflow integration, interpretability and trust when using the CAD system [13]. Implementation studies in India also show that AI tools are not just a product of model performance but are also influenced by the connectivity, training, equipment, and programme design of the tools [14].

Technologically, multimodal machine learning forms the basis for integrating images, audio, symptoms, and metadata into more powerful screening [15]. Despite all this, known pitfalls of TB computer-aided diagnosis-dataset shift, weak external validation, and limited domain adaptation-remain [16]. Infrastructure, hardware and quality control still need to be in place to ensure success; when they are, population-scale deployments illustrate the high throughput that automated radiographic screening can achieve [17].

Offline and edge processing offer a route to move screening intelligence from the cloud to the clinic, the mobile van, and community devices. Edge deployment on low-cost hardware indicates that on-device inference is feasible for radiographic TB screening [18]. Compact convolutional models likewise show that radiographic screening can be accurate while being engineered for clinical deployment under resource limitations [19].

This review examines automated multimodal TB screening systems operating under offline and edge constraints and organises them within a deployment-oriented classification. The contribution is fourfold: a classification of modalities, fusion strategies, deployment modes, automation functions, and human-machine interaction, consolidated from categories already established in the literature; a comparative analysis of representative systems; a deployment-readiness framework that operationalises these categories for low-resource settings; and a mapping of research gaps to a future-research roadmap. The overall system that frames this review is shown in Figure 1.

qq20260924-154305.png

Figure 1. Architecture of an automated multimodal tuberculosis (TB) screening system

TB screening is a triage procedure to identify people needing confirmatory testing; diagnosis is confirmed by microbiological identification of Mycobacterium tuberculosis. Screening is most useful in health care settings where confirmatory testing, radiology services, and specialist radiologists may be expensive, not local, or unavailable. Presumptive diagnosis can then be helpful in automated screening if the rate of false negatives and false positives is low.

The different types of systems can be distinguished by five axes: (1) modality, (2) the way the modalities are combined, (3) the deployment architecture of the system, (4) the automation function performed by the system, and (5) the way that humans interact with the system. Modality combinations specify the type of signals that the model is tuned to; the fusion strategy will describe the process used to combine signals; the deployment architecture will describe where inference is achieved; the automation function will indicate whether the system does triage, prioritisation, or referral support; the interaction mode will describe how the outputs are communicated to clinicians or community health workers. These five axes consolidate categories that are already well established in the multimodal-learning and edge-computing literature [21, 23]; they are used here as an organising classification for TB screening systems rather than as a claim of a new taxonomy.

For pulmonary TB, chest radiography is the most widely used modality, and lesions can be localised and evaluated automatically, making it suitable for a wide range of input modalities with only a small number of annotations, which can cause the loss of accuracy with image variability and lead to the development of hybrid transformer-convolutional neural network (CNN) architectures and customised loss functions [24]. In healthcare, multimodal learning is often beneficial as complementary data streams reinforce each other [25]. Cough sounds, symptoms, and clinical metadata provide accessible context, while AI imaging provides efficiency, not in lieu of clinical judgment [26]. Explainable models are required for practitioners to identify when they are confident, uncertain or when the quality of the data is poor [27] and confirmatory molecular and urine testing demonstrate that triage is enhanced when AI output is supported by practical confirmatory pathways, especially for people living with HIV [28]. Previous experience with broader respiratory AI indicates that validation, usability and reliability are key factors for the adoption of AI, rather than novelty [29]. The main inputs to the screening process, and the roles they typically automate, are summarised in Table 2; the confirmatory molecular pathway is listed only as the end-point to which screening must link, since it is a diagnostic rather than a screening modality.

Table 2. Overview of tuberculosis (TB) screening modalities and automation roles

Note: 1. The final row is a confirmatory diagnostic pathway rather than a screening modality; it is shown only as the end-point to which automated screening must link. 2. TB: tuberculosis; CXR: chest X-ray; AI: artificial intelligence; HIV: human immunodeficiency virus.

The most basic multimodal pairing is radiography and symptoms, as radiography gives visual cues and symptoms provide low-cost clinical context. Age, sex, HIV status, smoking history and previous TB treatment are known, and age and sex alone do not yield enough information for risk stratification, so radiography is used with the other information. Cough-and-symptom systems are useful when radiography is not available or is limited; more sophisticated systems include radiography, cough audio and notes for contextual screening [6].

The second design decision is the fusion strategy. Early fusion merges raw parts prior to representation learning, which might be unstable in case of modalities that are of different scale or quality. Intermediate fusion uses separate encoders and combines latent features, thus yielding richer cross-modal learning. Late fusion is a combination of per-modality scores and is better suited when the inputs are missing or low quality; adaptive fusion weights modalities based on their availability and quality. The fusion and deployment strategies are organised in Table 3.

Table 3. Classification of multimodal fusion and deployment strategies

The types of deployment architecture include cloud-assisted, hybrid edge-cloud, and offline edge. Cloud-assisted systems can handle large models and centralized auditing, but are susceptible to outages. Hybrid systems perform local screening and periodic training/auditing in the cloud. Offline systems maintain local inferences and work well in environments that have electricity, internet, and expert access issues [18]. A decision matrix comparing these three architectures is presented in Figure 3.

image005.png

Figure 3. Edge deployment decision matrix comparing architectures

Offline and edge deployment must be considered as a fundamental design criterion and not an added afterthought. Edge deployment of computer vision and medical diagnostic applications reveals limitations due to memory, energy, inference latency, model size, and hardware availability [50]. The clinical effectiveness and operational feasibility are thus important practical considerations for an edge model. The proposed deployment-readiness criteria are summarized in Table 5.

The first is latency. For mobile vans and screening sessions, a quick inference is needed in order to provide a triage evaluation during the patient visit. The second point is computational cost and footprint: A model that requires powerful graphics processing unit (GPU) servers may be attractive in benchmark terms but impractical in clinics that lack such resources and often face an unstable power supply. The third is privacy: on-device inference reduces the need to transfer chest X-rays, cough recordings and clinical data to remote servers.

These properties can be achieved by model compression. First, quantization lowers numerical precision so that the model uses less memory and less compute; second, pruning eliminates unnecessary parameters or channels; third, knowledge distillation moves capability from a large network to a smaller deployable one. Efficient convolutional networks can then be combined with device runtimes or hardware accelerators to realise fast inference.

Deployment readiness also depends on the specific hardware platform, because the same compressed model behaves differently across devices. Central processing unit (CPU)-only single-board computers such as the Raspberry Pi offer the lowest cost but the least computational headroom, so aggressive quantization is usually unavoidable; Jetson-class boards add an embedded GPU that supports larger models at a higher energy cost; and dedicated accelerators, such as the Ascend 310 chip used for radiographic TB screening [18] or FPGA-based pipelines [46], offer low latency within tight power budgets but tie the system to a specific toolchain. Quantization interacts with accuracy across all of these targets: post-training 8-bit quantization typically costs little accuracy when it is calibrated on representative data, but the degradation can grow under the distribution shift and class imbalance that are common in field screening. Deployment-ready reporting should therefore state accuracy, latency, memory and energy per hardware platform after compression, not only for the full-precision development model [21, 50].

The associated risks are also of significance. Domain shift is a decrease in performance after training on a different population or imaging device. Calibration risk exists because the predicted probability does not necessarily align with the true disease risk, which is why reporting of the Brier score, calibration curves and threshold transfer across sites is important. The missing-modality risk is when a cough is not present in the audio, metadata or imaging modality. The risks that arise in such areas as rural environments and mobile programmes are prevalent and are generally equipped with fallback strategies, per-modality confidence as well as quality indicators in deployment-ready systems.

qq20260924-152638.png

Figure 4. Field monitoring and model update loop

Regulatory clearance is a further readiness dimension that technical criteria alone do not capture. Commercially deployed radiographic CAD products illustrate the available pathways: CAD4TB and qXR are CE-marked medical devices under the European regulatory framework, qXR additionally holds United States Food and Drug Administration (FDA) 510(k) clearances for a set of chest X-ray findings, and the 2025 WHO policy statement identified six CAD products that meet WHO performance standards for TB screening in people aged 15 years and older [4]. National programmes increasingly require such recognition for procurement, and the Stop TB Partnership implementation guidance treats regulatory status as part of product selection [5]. By contrast, most research prototypes reviewed here, including the edge systems, report no clearance of any kind, which is precisely what separates a feasibility demonstration from a deployable medical device. Regulatory status is therefore included as an explicit criterion in Table 5.

Table 5. Deployment-readiness criteria for low-resource settings

Note: CXR: chest X-ray; AI: artificial intelligence; CE: European conformity marking; FDA: United States Food and Drug Administration; WHO: World Health Organization.

A deployment-ready system also contains a monitoring and update loop, as shown in Figure 4. In field use, drift detection can be made concrete by monitoring input and output distributions on the device: population-stability-index or Kolmogorov–Smirnov statistics computed on image-quality indices, cough signal-to-noise ratios and predicted-risk scores flag drift, while the abnormality-call rate and the proportion of AI-positive cases that are bacteriologically confirmed serve as label-free performance proxies. Checks of this kind can run automatically over a rolling window of recent cases, with alert limits fixed at commissioning; an exceedance triggers a threshold review in which the operating point is re-derived from a recent, locally verified sample. Any model update is then validated locally against a fixed site-specific reference set – several hundred consecutive cases with known confirmatory outcomes is a practical minimum – before release, and the previous version is restored if sensitivity, specificity or calibration degrades. The main contribution of this review to system engineering is the transformation of this static model into an automated screening system that can be maintained safely over time.