Andrew K. C. Wong

The healthcare industry seeks to integrate AI into clinical applications, yet understanding AI decision making remains a challenge for healthcare practitioners as these systems often function as black boxes. Our work benchmarks the Pattern Discovery and Disentanglement (PDD) system’s unsupervised learning algorithm, which provides interpretable outputs and clustering results from clinical notes to aid decision making. Using the MIMIC-IV dataset, we process free-text clinical notes and ICD-9 codes with Term Frequency-Inverse Document Frequency and Topic Modeling. The PDD algorithm discretizes numerical features into event-based features, discovers association patterns from a disentangled statistical feature value association space, and clusters clinical records. The output is an interpretable knowledge base linking knowledge, patterns, and data to support decision making. Despite being unsupervised, PDD demonstrated performance comparable to supervised deep learning models, validating its clustering ability and knowledge representation. We benchmark interpretability techniques—Feature Permutation, Gradient SHAP, and Integrated Gradients—on the best-performing models (in terms of F1, ROC AUC, balanced accuracy, etc.), evaluating these based on sufficiency, comprehensiveness, and sensitivity metrics. Our findings highlight the limitations of feature importance ranking and post hoc analysis for clinical diagnosis. Meanwhile, PDD’s global interpretability effectively compensates for these issues, helping healthcare practitioners understand the decision-making process and providing suggestive clusters of diseases to assist their diagnosis.

Medical datasets may be imbalanced and contain errors due to subjective test results and clinical variability. The poor quality of original data affects classification accuracy and reliability. Hence, detecting abnormal samples in the dataset can help clinicians make better decisions. In this study, we propose an unsupervised error detection method using patterns discovered by the Pattern Discovery and Disentanglement (PDD) model, developed in our earlier work. Applied to the large data, the eICU Collaborative Research Database for sepsis risk assessment, the proposed algorithm can effectively discover statistically significant association patterns, generate an interpretable knowledge base for interpretability, cluster samples in an unsupervised learning manner, and detect abnormal samples from the dataset. As shown in the experimental result, our method outperformed K-Means by 38% on the full dataset and 47% on the reduced dataset for unsupervised clustering. Multiple supervised classifiers improve accuracy by an average of 4% after removing abnormal samples by the proposed error detection approach. Therefore, the proposed algorithm provides a robust and practical solution for unsupervised clustering and error detection in healthcare data.

Background: Statistical data analysis, especially the advanced machine learning (ML) methods, have attracted considerable interest in clinical practices. We are looking for interpretability of the diagnostic/prognostic results that will bring confidence to doctors, patients and their relatives in therapeutics and clinical practice. When datasets are imbalanced in diagnostic categories, we notice that the ordinary ML methods might produce results overwhelmed by the majority classes diminishing prediction accuracy. Hence, it needs methods that could produce explicit transparent and interpretable results in decision-making, without sacrificing accuracy, even for data with imbalanced groups. Methods: In order to interpret the clinical patterns and conduct diagnostic prediction of patients with high accuracy, we develop a novel method, Pattern Discovery and Disentanglement for Clinical Data Analysis (cPDD), which is able to discover patterns (correlated traits/indicants) and use them to classify clinical data even if the class distribution is imbalanced. In the most general setting, a relational dataset is a large table such that each column represents an attribute (trait/indicant), and each row contains a set of attribute values (AVs) of an entity (patient). Compared to the existing pattern discovery approaches, cPDD can discover a small succinct set of statistically significant high-order patterns from clinical data for interpreting and predicting the disease class of the patients even with groups small and rare. Results: Experiments on synthetic and thoracic clinical dataset showed that cPDD can 1) discover a smaller set of succinct significant patterns compared to other existing pattern discovery methods; 2) allow the users to interpret succinct sets of patterns coming from uncorrelated sources, even the groups are rare/small; and 3) obtain better performance in prediction compared to other interpretable classification approaches. Conclusions: In conclusion, cPDD discovers fewer patterns with greater comprehensive coverage to improve the interpretability of patterns discovered. Experimental results on synthetic data validated that cPDD discovers all patterns implanted in the data, displays them precisely and succinctly with statistical support for interpretation and prediction, a capability which the traditional ML methods lack. The success of cPDD as a novel interpretable method in solving the imbalanced class problem shows its great potential to clinical data analysis for years to come.

A protein family has similar and diverse functions locally conserved. An aligned pattern cluster (APC) can reflect the conserved functionality. Discovering aligned residue associations (ARAs) in APCs can reveal subtle inner working characteristics of conserved regions of protein families. However, ARAs corresponding to different functionalities/subgroups/classes could be entangled because of subtle multiple entwined factors. To discover and disentangle patterns from mixed-mode datasets, such as APCs when the residues are replaced by their fundamental biochemical properties list, this paper presents a novel method, Extended Aligned Residual Association Discovery and Disentanglement (E-ARADD). E-ARADD discretizes the numerical dataset to transform the mixed-mode dataset into an event-value dataset, constructs an ARA Frequency Matrix and then converts it into an adjusted Statistical Residual (SR) Vector Space (SRV) capturing statistical deviation from randomness. By applying Principal Component (PC) Decomposition on SRV, PCs ranked by their variance are obtained. Finally, the disentangled ARAs are discovered when the projections on a PC is re-projected to a vector space with the same basis vectors of SRV. Experiments on synthetic, cytochrome c and class A scavenger data have shown that E-ARADD can a) disentangle the entwined ARAs in APCs (with residues or biochemical properties), b) reveal subtle AR clusters relating to classes, subtle subgroups or specific functionalities. E-ARADD can discover and disentangle ARs and ARAs entangled in functionality and location of protein families to reveal functional subgroups and subgroup characteristics of biological conserved regions. Experimental results on synthetic data provides the proof-of-concept validation on the successful disentanglement that reveals class-associated ARAs with or without class labels as input. Experiments on cytochrome c data proved the efficacy of E-ARADD in handing both types of residue data. Our novel methodology is not only able to discover and disentangle ARs and ARAs in specific statistical/functional (PCs and RSRVs) spaces, but also their locations in the protein family functional domains. The success of E-ARADD shows its great potential to proteomic research, drug discovery and precision and personalized genetic medicine.

Residue-residue close contact (R2R-C) data procured from three-dimensional protein-protein interaction (PPI) experiments is currently used for predicting residue-residue interaction (R2R-I) in PPI. However, due to complex physiochemical environments, R2R-I incidences, facilitated by multiple factors, are usually entangled in the source environment and masked in the acquired data. Here we present a novel method, P2K (Pattern to Knowledge), to disentangle R2R-I patterns and render much succinct discriminative information expressed in different specific R2R-I statistical/functional spaces. Since such knowledge is not visible in the data acquired, we refer to it as deep knowledge. Leveraging the deep knowledge discovered to construct machine learning models for sequence-based R2R-I prediction, without trial-and-error combination of the features over external knowledge of sequences, our R2R-I predictor was validated for its effectiveness under stringent leave-one-complex-out-alone cross-validation in a benchmark dataset, and was surprisingly demonstrated to perform better than an existing sequence-based R2R-I predictor by 28% (p: 1.9E-08). P2K is accessible via our web server on https://p2k.uwaterloo.ca.

Functional region identification is of fundamental importance for protein sequences analysis. Such knowledge provides better scientific understanding and could assist drug discovery. Up-to-date, domain annotation is one approach but it needs to leverage existing databases. For de novo discovery, motif discovery locates and aligns locally homologous sub-sequences to obtain a position-weight matrix (PWM) which is a fixedlength representation model whereas protein functional region size varies. It thus requires computational expensive exhaustive search to obtain a PWM with width of optimal range. This paper presents a new method known as Pattern-Directed Aligned Pattern Clustering (PD-APCn) to discover and align patterns in conserved protein functional regions. It adopts Aligned Pattern Cluster (APC) with patterns of variable length and strong support to direct the incremental APC expansion. It allows substitution and frame-shift mutations until a robust termination condition is reached. The concept of breakpoint gap is introduced to identify spots of mutations such as substitution and frameshifts. Experiments on synthetic datasets with different size and noise level showed that PD-APCn outperforms MEME with much higher recall and Fmeasure and computational speed 665 times faster that MEME. When applying to Cytochrome C and Ubiquitin families, it found all key binding sites within the APCs.

A protein family has similar and diverse functions locally conserved as aligned sequence segments. Further discovering their association patterns could reveal subtle family subgroup characteristics. Since aligned residues associations (ARAs) in Aligned Pattern Clusters (APCs) are complex and intertwined due to entangled function, factors, and variance in the source environment, we have recently developed a novel method: Aligned Residue Association Discovery and Disentanglement (ARADD) to solve this problem. ARADD first obtains from an APC an ARA Frequency Matrix and converts it to an adjusted statistical residual vector space (SRV). It then disentangles the SRV into Principal Components (PCs) and Re-projects their vectors to a SRV to reveal succinct orthogonal AR groups. In this study, we applied ARADD to class A scavenger receptors (SR-A), a subclass of a diverse protein family binding to modified lipoproteins with diverse biological functionalities not explicitly known. Our experimental results demonstrated that ARADD can unveil subtle subgroups in sequence segments with diverse functionality and highly variable sequence lengths. We also demonstrated that the ARAs captured in a Position Weight Matrix or an APC were entangled in biological function and domain location but disentangled by ARADD to reveal different subclasses without knowing their actual occurrence positions.