A Dual-Memory Cognitive Algorithm for Heart Disease Diagnosis: An Unconscious Mind-Inspired Model with Pattern Consolidation and k-Nearest Neighbor Retrieval

Benchmark studies on the Cleveland Heart Disease dataset have reported strong performance across several machine learning models. Some studies have shown that gradient boosting methods can achieve accuracies in the range of 85% to 88%. In addition, instance-based approaches such as weighted k-nearest neighbor have achieved about 84.2% accuracy when the number of neighbors was set between 5 and 9.

Recent work on biomedical optimization and machine learning includes the development of deep learning algorithms for COVID-19 classification from chest X-rays, achieving high accuracy with the VGG-16 architecture [22]. As well as Acharya et al. [23] demonstrated stochastic mathematical modeling with optimal control for type 2 diabetes management. These works demonstrate systematic optimization approaches for biomedical classification tasks.

The transition of medical diagnosis using machine learning techniques has been remarkably dynamic and dynamic within the last four decades, ranging from traditional rule-based expert systems to developing more complex deep learning models that can handle multiple sources of medical data. Although early medical artificial intelligence tools, dating back to the 1970s and 1980s, relied on expressing domain expertise as logical rules that can be directly processed and implemented using inference engines, as seen in some pioneering projects like MYCIN (an early expert system for infectious disease diagnosis) for bacterial infections and INTERNIST (a computer-assisted diagnostic system for internal medicine) for general internal medicine [14], they showed some limitations and challenges, including rigidity for unfamiliar instances, knowledge acquisition and maintenance, and learning from experience [15]. These limitations and challenges have led researchers and scientists to look for an alternative paradigm shift that uses data-driven approaches and machine learning algorithms capable of automatically discovering medical diagnostic knowledge from datasets without relying on rules [24].

The emergence of statistical learning theory in the 1990s enabled a mathematical framework for machine learning and spawned the concept of SVMs that enable optimal identification of decision boundaries within high-dimensional feature space using kernel techniques [17]. SVMs have been widely implemented for CVD risk prediction with multiple publications proving their efficacy for maximum-margin hyperplane determination with optimal discrimination between disease and non-disease classes [25]. The ‘kernel trick’ allows SVMs to effectively operate within higher-dimensional feature space and tackle the non-linear relationship existing between clinical variables and disease presence [26]. Nonetheless, the determination of optimal kernel functions and hyperplanes still presents complexities with SVMs and optimal decision boundaries that are more abstract from an intuitive interpretation perspective [27].

Cognitive-inspired computing is a paradigm shift within artificial intelligence research that focuses on mimicking and understanding the information processing architectures within biological brains instead of viewing intelligence as an optimization task [16]. It is recognized within cognitive computing that human intelligence arises out of highly complex neural systems, which have evolved as an optimized solution within biological constraints for solving problems effectively within the world [27]. Examples within early cognitive computing include perceptrons and multi-layer connections inspired by biologically modeled connections within biological nerves, but these were very loosely based on biologically modeled connections and were nowhere near any realistic modeling of brain functions [28]. Contemporary advances within the realms of cognitive neuroscience have enabled more realistic modeling and implementation within cognitive architectures of working memory buffers and more [29].

Memory-augmented neural networks are an important step forward with regard to cognitive architectures and deep learning, as they incorporate specific external memory components that may be selectively read and written via attention operations [30]. The Neural Turing Machine, originally introduced by Graves et al. [31] as an extension to Turing machines, uses a neural network controller and an external memory that can be accessed via read and write operations. Differential neural computers extend these capabilities with more advanced addressing of memory components, including lookup operations based on content and links over time [32]. Memory networks, introduced as a solution for question answering tasks with separate memory components that have an attention mechanism, were also introduced by Weston et al. [33]. These networks and architectures clearly show that cognitive architectures improve the capabilities and skills of neural networks that have to reason about memory and achieve these goals at an abstract level distant from the cognitive abilities of humans [34].

The idea of unconscious mental processes has an illustrious academic background that goes as far back as philosophical and psychoanalytic interpretations and continues with cognitive and neuroscientific theories. Several philosophers [35] realized that conscious processing constitutes a very small percentage of mental functions, and an immense amount goes on unconsciously. Nevertheless, it was Sigmund Freud who introduced the first theoretical concept related to the unconscious mind, suggesting that suppressed thoughts, desires, and conflicts still impact behavior despite being beyond conscious access [28]. Although several Freudian theories have been proven wrong or refuted by modern science, Freud's observation about the relevance of unconscious processing has been vindicated [36].

Freud's structural model consists of three parts: the id, ego, and superego. The id represents the unconscious instinctual impulses, the ego represents the conscious logical side of personality, and the superego represents internalized morality [26]. The ego uses defense mechanisms like repression to suppress threatening ideas from becoming conscious, but these have an indirect effect on consciousness via symptoms, dreams, and slips of the tongue [37]. Although modern cognitive science does not subscribe to Freudian theory on these points, concepts like selective filtering of information, division between accessible and inaccessible memories, and an influence of previous experiences on current processing without consciousness have been proven empirically [38]. Simulation of defense mechanisms and unconscious processing can be seen as a metaphorical interpretation of Freudian ideas and not as an actual implementation [39].

The optimal UMIA reached an accuracy of 86.67% on the 60-case testing set, which was equal to that of Random Forest and superior to all remaining traditional machine learning methods. The precision achieved by UMIA was 87.22%, surpassing that of Random Forest 86.77%, SVM 85.26%, XGBoost 85.26%, Logistic Regression 83.41%, and LightGBM 81.00%. The three best methods have equal recall at 86.67%, and LightGBM performed poorly at 80.00% recall. The F1-scores varied from 0.8655 for UMIA to 0.7966 for LightGBM.

Both SVM and XGBoost converged on the same confusion matrix with 29 true negatives (TNs), 3 false positives (FPs), 22 true positives (TPs), and 6 false negatives (FNs), implying they have identified an equivalent boundary despite their divergent optimization philosophies. The difference of 6.67 percentage points from 80% to 86.67% accuracy marks an essential difference, but validation via statistical testing would be necessary for verification.

The confusion matrix obtained from UMIA yielded 30 instances of true negatives and 2 instances of false positives among 32 actual negatives, thereby giving a specificity value of 93.75%, while there were 22 instances of true positives and 6 instances of false negatives among 28 actual positives, giving a sensitivity value of 78.57%. The model had a positive predictive value and a negative predictive value of 92% and 83%, respectively. It shows bias towards specificity and not towards sensitivity.

This work has applied formal statistical tests to determine whether the observed performance differences among the six models reflected true algorithmic advantages or were simply due to random variation by using a Friedman test, which is appropriate for comparing multiple related models. The developed system has found no significant differences in performance χ² = 3.47, p = 0.63, df = 5. Because this p-value is far above 0.05, and it could not reject the assumption that all models perform similarly, given the relatively small test set of 60 samples, the observed accuracy range of 80.0% to 86.7% is consistent with normal sampling variability rather than meaningful superiority.

Pairwise paired t-tests between the UMIA and each competing method led to the same conclusion, and for comparisons with Random Forest, SVM, and LightGBM, all produced non-significant results, all p > 0.48, and the remaining comparisons showed similarly high p-values. Together, these findings indicate that UMIA matches the performance of established models rather than outperforming them.

Accordingly, this work could frame UMIA’s contribution around interpretability rather than accuracy gains, as UMIA delivers competitive diagnostic performance while offering transparency features unavailable in black-box approaches, including access to 56 interpretable memory patterns and retrieval of seven similar cases with quantified similarity calibrated confidence scores and frequency-based pattern importance. In clinical decision-support settings where trust and explainability are critical, such transparency can justify adoption even in the absence of statistically superior accuracy.

Figure 5 compares the predictive performance of UMIA and the benchmark models across the main evaluation metrics, showing that UMIA remained competitive with the best-performing conventional classifiers.

Figure 5. Comprehensive model performance comparison

Figure 6. Confusion matrix: Unconscious Mind-Inspired Algorithm (UMIA) optimized

Figure 6 presents the confusion matrix of the optimized UMIA model, highlighting its strong specificity and its tendency to produce fewer false positives than false negatives.

Random Forest performed slightly better on sensitivity (sensitivity: 82.14% vs. 78.57%, with 5 vs. 6 false negatives), but with less specificity (specificity: 90.63% vs. 93.75%, with 3 vs. 2 false positives). SVM and XGBoost were comparable with UMIA in sensitivity but less specific. LightGBM performed poorly due to 9 false negatives, which were about one-third of all disease instances, and was too conservative on positive calls. Logistic Regression had the most lenient threshold with 4 false positives and 6 false negatives.

Figure 7 compares the confusion matrices of all evaluated models and shows the trade-off each method makes between sensitivity and specificity after learning that there were 56 patterns stored within conscious memory, reflecting 37.3% of capacity, which represents selective storage as opposed to complete retention. The unconscious buffer contained 181 patterns waiting on consolidation. The 237 learning instances were all added to the associative weights dictionary, representing complete retention despite selective storage. The average pattern frequency value of 1.00 with a maximum value of 1.00 shows very limited pattern fusion, reflecting a level of diversity among learning instances such that very few, if any, reached the 0.70 similarity threshold for fusion.

Figure 7. Confusion matrix of all models

Figure 8. Unconscious Mind-Inspired Algorithm (UMIA) confidence scores distribution

Although there is 37.3% memory usage, there is still room for more learning without encountering the problem of catastrophic forgetting. A large number of unconscious samples, 181 samples, show that most samples have a high degree of similarity with existing memories and thus are redundant. Within a continual learning scenario, some unconscious samples might have gathered enough instances to be promoted as a result of repeated presentations. Figure 8 shows the distribution of UMIA confidence scores on the test set, indicating that most predictions were made with moderate to high confidence and that only a small fraction of cases fell into the uncertain range.

Confidence values varied from 0.45 to 0.73, with a focus on 0.60 and 0.65, suggesting that, on average, there were fairly good matches made within conscious memory for most testing instances. About 40% of instances showed confidence above 0.65, indicating high certainty, and just 5% below 0.55, indicating uncertain classifications. The fact that no confidence values were near 1.0 supports that no testing instance similarity existed with instances within the train set, and correspondingly, values below 0.45 suggest even unusual testing instances were similar.

The Friedman test obtained a chi-squared value of 3.4694 with a p-value of 0.6280, which failed to reject the null hypothesis that there were equal performances among all algorithms. The result implies that it was not unlikely, given the small sample size of 60 instances, that accuracy variability among 80%-86.67% could occur due to chance. The result of the paired t-test on UMIA compared with all competitors on performance showed insignificant values with p-values ranging from 0.32 to 0.74. The UMIA vs. Random Forest t-statistic value was -0.7041 with p-value 0.4841, UMIA vs. SVM t-statistic value -0.5741 with p-value 0.5681, and UMIA vs. LightGBM t-statistic value 0.6293 with p-value 0.5316, indicating that it might perform better.

Such observations indicate that several algorithms perform similarly on this data set and that algorithm criteria are leaning more toward applicability and relevance instead of purely accuracy measures.

Comparison between optimized UMIA with 86.67% accuracy and baseline implementation with 68.33% accuracy shows absolute improvement of 18.34 percentage points, which signifies 26.84% relative improvement and 57.9% error reduction. So, it clearly requires emphasis on systematic parameter adjustment for cognitive-inspired models. The baseline set with conservative parameters (memory size = 100, association strength = 0.10, connection threshold = 0.50, learning rate = 0.01) caused sparse storage, low pattern transformation, and sluggish association learning. These deficiencies were remedied with enlarged capacity, superior pattern transformation, more flexible admission criteria, and faster learning.

The training time of 2.34 seconds for UMIA stands between fast approaches, SVM 0.12 s, Logistic Regression 0.08 s, and slower ensemble learning methods, XGBoost 1.87s, LightGBM 0.98 s, while Random Forest took 0.45 seconds. The prediction time of 8.3 milliseconds per sample is higher compared to competing approaches, 1.5–4.2 ms, but still optimal for real-time processing. The memory usage of 1.2 MB is relatively low compared to ensemble learning methods, Random Forest 8.5 MB, XGBoost 5.3 MB, and LightGBM 4.1 MB, due to UMIA learning and selecting 56 patterns instead of the whole datasets or large ensembles. Its low memory usage is highly advantageous.

This work analyzed the model’s misclassified cases to understand why errors occurred. The unconscious mind-inspired algorithm produced six false negatives, 21.4% of true disease cases, and two false positives, 6.3% of true non-disease cases, and compared these cases with correctly classified patients using simple statistical tests.

False negatives stood out mainly because they achieved much higher exercise heart rates than correctly detected patients, indicating preserved exercise capacity despite disease. Clinically, this suggests early or atypical disease that does not yet limit performance, making detection harder. Other differences, such as age, cholesterol, and segment depression, followed expected trends but were not statistically meaningful given the small number of cases.

A detailed quantitative comparison between false negatives and true positives is presented in Table 2, which helps identify the clinical characteristics associated with missed disease cases. A corresponding comparison between false positives and true negatives is presented in Table 3, which helps explain why some healthy cases were incorrectly classified as diseased.

Error analysis reveals several courses of action based on error characteristics, these including incorporating timing information about disease progression to separate early disease from healthy subjects, incorporating variables reflecting disease severity or loss of function, using more complex measures of similarity based on vector weights specified based on relevance to diagnosis, using confidence thresholds indicating predictions as indeterminate if below a threshold, and acquiring more data for difficult populations.

Table 2. Quantitative error analysis: Feature comparison for misclassified cases, false negatives (n = 6) vs. true positives (n = 22)

Table 3. Quantitative error analysis: Feature comparison for misclassified cases, false positives (n = 2) vs. true negatives (n = 30)

Figure 9. Correlation heat map

Figure 9 shows prominent correlations among various clinical parameters and disease status. Strongest positive correlations include thalassemia (0.53), chest pain type (0.41), and ST depression (0.42). Max heart rate has a negative correlation (−0.42), as would be expected with decreased exercise tolerance.

This work performed a systematic sensitivity analysis to empirically justify the selection of k = 7. Seven different values of k 3, 5, 7, 9, 11, 13, and 15 were evaluated while keeping all other UMIA parameters fixed: memory size = 150, associated strength is 0.15, connection threshold = 0.4, and learning rate = 0.05, and each configuration was assessed using the same 60-sample test set and the full results are summarized in Table 1.

The analysis highlights clear tradeoffs associated with different choices of k; smaller values k = 3 and k = 5 showed moderate sensitivity, 57.14% and 53.57%, respectively, but were associated with higher false-positive counts, four and three false positives. In contrast, k = 7 achieved the best overall balance, yielding the highest F1-score 0.7415 and an accuracy of 75.0%. The complete results of the sensitivity analysis for different values of k are summarized in Table 4, which shows that k = 7 achieved the best overall balance between accuracy, F1-score, sensitivity, and specificity.

Table 4. k-neighbors sensitivity analysis results (n = 60 test samples)

Table 5. Explainability comparison

Decision trees have high rule transparency with no confidence scoring or no case retrieval. Logistic Regression has global feature importance as no instance explanations while standard k-NN either basic case retrieval and no memory architecture or no importance weighting but UMIA enhancements firstly dual-memory inspection 56 conscious, 181 unconscious then confidence scoring mean similarity across k = 7 also pattern importance novelty + frequency weighting lastly similarity metrics 70% cosine + 30% Euclidean and Table 5 compares five explainability dimensions across models.