A hierarchical machine learning model for predicting self-harm and suicidal behaviour in hospitalised patients with schizophrenia using clinical history and nursing observations.

Objective: This study aimed to develop and evaluate a two-layered machine learning framework that combines admission clinical information with longitudinal nursing observations to identify schizophrenia inpatients at high risk of self-harm or suicidal acts. Methods: We retrospectively reviewed the records of 477 patients with schizophrenia hospitalised in Liaoning Province between July 2021 and July 2024. According to whether at least one self-injurious or suicidal episode was documented during the index admission, 159 individuals were assigned to a high-risk group and 318 to a non-high-risk group. At admission, 18 baseline variables (including age, sex, history of self-harm, hopelessness/depression, and educational attainment) were extracted from electronic medical records, and 39 nurse-rated behavioural items were scored weekly using the Psychiatric Patient Nursing Observation Scale. Static and dynamic feature sets were used to train six classifiers [regularized logistic regression (LR), support vector machine (SVM), extreme gradient boosting, random forest, multi-layer perceptron, and K-nearest neighbours]. The best static model (regularized LR) and the best dynamic model (SVM) were combined through probability-level weighted fusion to generate a hierarchical risk score. Results: Multivariable analysis of admission features showed that previous self-harm [odds ratio (OR) = 4.323], hopelessness/depression (OR = 3.090), younger age (OR = 0.938), and higher educational level (OR = 1.357) were independent predictors of self-harm/suicidal behaviour. Among dynamic indicators, negative self-evaluation (OR = 2.303), self-reported depression (OR = 1.812), insomnia (OR = 1.768), talking to oneself (OR = 1.733), crying (OR = 1.700), and reduced conversation with others (OR = 1.422) remained significant. The optimised static LR model achieved an area under the curve (AUC) of 0.7564, and the dynamic SVM model reached an AUC of 0.8531. Their fusion further improved performance (AUC = 0.9048; sensitivity 0.8542; specificity 0.7789; accuracy 0.8042). This hierarchical model outperformed the best flat combined-feature model (SVM; AUC = 0.9022) in sensitivity (0.8542 vs. 0.6667), indicating a more clinically appropriate detection of high-risk patients. Conclusion: A hierarchical machine learning approach that integrates baseline clinical history with repeated nursing assessments can effectively flag schizophrenia inpatients at high risk for self-harm and suicidal behaviour, supporting timely and individualised preventive strategies in psychiatric wards.