Transdiagnostic clustering and network analysis for questionnaire-based symptom profiling and drug recommendation in the UK Biobank and a Korean cohort

Abstract

Clinical decision support systems (CDSSs) play a critical role in enhancing the efficiency of mental health care delivery and promoting patient engagement. Transdiagnostic approaches that utilize raw psychological and biological data enable personalized patient profiling and treatment. This study introduces a CDSS incorporating symptom profiling and drug recommendation for mental health care. Among the UK Biobank cohort, we analyzed 157,348 participants for symptom profiling and 14,358 participants with a drug prescription history for drug recommendation. Among the 1307 patients in the Samsung Medical Center cohort, 842 were eligible for analysis. Symptom profiling utilized demographic and questionnaire data, employing conventional clustering and community detection methods. Identified clusters were explored using diagnostic mapping, feature importance, and scoring. For drug recommendation, we employed cluster- and network-based approaches. The analysis identified nine clusters using k-means clustering and ten clusters with the Louvain method. Clusters were annotated for distinct features related to depression, anxiety, psychosis, drug addiction, and self-harm. For drug recommendation, drug prescription probabilities were retrieved for each cluster. A recommended list of drugs, including antidepressants, antipsychotics, mood stabilizers, and sedative-hypnotics, was provided to individual patients. This CDSS holds promise for efficient personalized mental health care and requires further validation and refinement with larger datasets, serving as a valuable tool for mental healthcare providers.

Figure 1

Illustration of information regarding diagnosis and results from symptom profiling based on data from 157,348 subjects using UMAP (Uniform Manifold Approximation and Projection). (A) Self-reported diagnosis. (B) Symptom-based diagnosis. (C) Result of k-means clustering. (D) Result of community detection applying the Louvain algorithm.

Figure 2

Analysis for cluster identification. (A) Comparison across clustering results, self-reported Dx and symptom-based Dx. Comparison between KM clustering results and Dx information (Left). Comparison between LV community detection results and Dx information (Middle). Comparison between KM clustering results and LV community detection results (Right). (B) Key questions identified in each cluster based on logistic regression. (C) Distribution of scores for the question categories in each cluster.

Figure 3

Statistics for cluster-based drug recommendation. (A) The percentile of samples with a history of psychiatric drugs for each cluster. The overall prescription rate across all clusters was 9.12% (red line). (B) The probability of prescribing each drug class to patients in a cluster.

Figure 4

Example of drug recommendations for patients with similar symptom profiles. Cluster-based drug recommendation (Left). All samples with similar symptom profile are more likely to be assigned to the same cluster as a result of symptom profiling, such that the same list of drugs will be recommended for all samples. Network-based drug recommendation (Right). In spite of similar symptom profiles, each sample has a network of neighbors with different drug prescription histories, and a different drug list will be recommended for each sample.

Figure 5

Analysis of the Korean cohort. (A) Illustration of information on diagnosis and results from symptom profiling of 842 patients using UMAP. (From left to right) Primary diagnosis; Scale-based diagnosis; Result of k-means clustering; Result of community detection applying Louvain algorithm. (B) Comparison across clustering results, Primary Dx and Scale-based Dx. (C) Key questions in each cluster as a result of logistic regression. *Abbreviations. CRS: Clinical Rating Scale; HRS: Hamilton Rating Scale; SR: Self-report; APPQ: Albany Panic and Phobia Questionnaire; ASI: Anxiety Sensitivity Index-3; BAI: Beck Anxiety Inventory; BDI: Beck Depression Inventory-II; MINI: Structured Interview; MDE: Major Depressive Episode. (D) Distribution of scores for the question categories in each cluster. (E) The probability of prescribing each drug class to patients in a cluster.