New technology for multi-class psychiatric disorder classification: An interpretable ensemble machine learning framework using actigraphy and demographic data
NeuroMarkers · 2026-11 · 100232
DOI: 10.1016/j.neumar.2026.100232 ↗Kurz gefasst
Fachartikel in der Zeitschrift NeuroMarkers, 2026-11. Verfasst von Suyaib Al Sifat und 2 weiteren Autor:innen. Die Arbeit nennt 72 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Objectives
Accurate diagnosis of psychiatric disorders remains a critical challenge, as traditional methods often rely on subjective evaluations that may delay or misguide treatment. There is a pressing need for objective and data-driven solutions to deliver timely and reproducible psychiatric screening. This study used actigraphy data and demographic metadata to develop a modular machine learning architecture for the classification of healthy controls, major depressive disorder, and schizophrenia.
Methods
A feature space of 24 variables was engineered across five clinical domains: circadian stability, sleep and inactivity structure, statistical distribution, physical intensity, and demographics. To achieve strong generalizability, eleven classifiers were trained using a leave-one-subject-out cross-validation protocol, and their hyperparameters were fine-tuned using a grid search.
Results
Soft voting and stacking architectures further improved the performance, with the top ensemble model achieving an accuracy of 89.6% and a macro-averaged area under the receiver operating characteristic curve of 0.94. In assessing model interpretation, SHapley Additive exPlanations values, permutation importance, and native feature rankings were used as reliable indicators of clinically relevant predictors. In addition, an ablation study across the five clinical domains confirmed that sleep architecture and circadian stability provided the highest diagnostic gain. The accuracy and reliability of the models were supported by reliability diagrams, temperature scaling, expected calibration error, Brier scores, and the exact McNemar test with Holm-Bonferroni correction.
Conclusion
Unlike prior applications of machine learning in psychiatry, which were generally restricted to binary classification, single-modality inputs, or ad hoc pipelines, our framework addresses these critical gaps. It integrates multi-class disorder coverage, actigraphy records enhanced with demographic metadata, domain-specific feature engineering, ensemble strategies, interpretability methods, and calibration testing in a standardized architecture. The proposed framework provides a scalable, interpretable, and transparent approach for automated psychiatric screening and digital mental health applications.
Zitieren
APA
Sifat, S. A., Ruba, S. A., & Uddin, K. M. M. (2026). New technology for multi-class psychiatric disorder classification: An interpretable ensemble machine learning framework using actigraphy and demographic data. NeuroMarkers, 100232. https://doi.org/10.1016/j.neumar.2026.100232
Vancouver
Sifat SA, Ruba SA, Uddin KMM. New technology for multi-class psychiatric disorder classification: An interpretable ensemble machine learning framework using actigraphy and demographic data. NeuroMarkers. 2026:100232. doi:10.1016/j.neumar.2026.100232
BibTeX
@article{Sifat2026,
author = {Sifat, Suyaib Al and Ruba, Sadia Amin and Uddin, Khandaker Mohammad Mohi},
title = {New technology for multi-class psychiatric disorder classification: An interpretable ensemble machine learning framework using actigraphy and demographic data},
journal = {NeuroMarkers},
year = {2026},
pages = {100232},
doi = {10.1016/j.neumar.2026.100232}
}
EUVIMED