Beyond Diagnosis: Data-Driven Psychiatric Subtypes for Precision Mental Health Care in an Underrepresented African Population

Background: Symptoms overlap and heterogeneity observed in psychiatric disorders pose a challenge to accurate diagnosis and patient stratification. In low-resource settings such as Nigeria, utilizing available structured clinical data with unsupervised machine learning can reveal latent subtypes, potentially improving diagnosis and treatment. Aims: This study aims to identify meaningful psychiatric subtypes in a Nigerian cohort using clustering of mental state and patient history data. Method: Data from 664 patients with 38 clinical variables were analyzed. Mental state domain included mood, thought content, perception, and insight, while history domain covered psychosocial background, family history, substance use, and past psychiatric episodes. After preprocessing, imputation and one-hot encoding, KMeans clustering was applied independently to the two domains. The optimal cluster number (k = 4) was chosen based on the Elbow Method and silhouette scores for clinical interpretability and model robustness. Cluster stability was assessed with Adjusted Rand Index and Normalized Mutual Information. Chi-square tests evaluated cluster-diagnosis correlation. Results: Mental state clustering demonstrated high stability (ARI = 0.9944; NMI = 0.9912), while history clustering showed moderate agreement (ARI = 0.4490; NMI = 0.5465). Correlation analysis between clusters identified four patient profiles with distinct clinical features: (1) a Social Stressor profile marked by reactive symptoms to acute social triggers without prior history, likely benefiting from psychosocial support; (2) a Psychiatric Chronicity profile of episodic or long-term disorders requiring ongoing management; (3) a Minimal Risk Factor profile representing mostly first-episode cases without clear risk factors, needing careful monitoring; and (4) a Complex Multi-Risk profile involving overlapping personal, familial, and medical risks requiring comprehensive multidisciplinary care. These profiles were significantly associated with diagnosis (p < 0.001). Conclusions:Unsupervised clustering revealed clinically relevant psychiatric subtypes in an underrepresented population, highlighting opportunities for tailored interventions. Further studies should validate these findings and explore their predictive value for treatment outcomes.