TY - JOUR
AU - J. Eder
AU - M. S. Dong
AU - M. Wöhler
AU - M. S. Simon
AU - C. Glocker
AU - L. Pfeiffer
AU - R. Gaus
AU - J. Wolf
AU - K. Mestan
AU - H. Krcmar
AU - N. Koutsouleris
AU - A. Schneider
AU - J. Gensichen
AU - R. Musil
AU - P. Falkai
A1 - 
AB - General practitioners play an essential role in identifying depression and are often the first point of contact for patients. Current diagnostic tools, such as the Patient Health Questionnaire-9, provide initial screening but might lead to false positives. To address this, we developed a two-step machine learning model called Clinical 15, trained on a cohort of 581 participants using a nested cross-validation framework. The model integrates self-reported data from validated questionnaires within a study sample of patients presenting to general practitioners. Clinical 15 demonstrated a balanced accuracy of 88.2% and incorporates a traffic light system: green for healthy, red for depression, and yellow for uncertain cases. Gaussian mixture model clustering identified four depression subtypes, including an Immuno-Metabolic cluster characterized by obesity, low-grade inflammation, autonomic nervous system dysregulation, and reduced physical activity. The Clinical 15 algorithm identified all patients within the immuno-metabolic cluster as depressed, although 22.2% (30.8% across the whole dataset) were categorized as uncertain, leading to a yellow traffic light. The biological characterization of patients and monitoring of their clinical course may be used for differential risk stratification in the future. In conclusion, the Clinical 15 model provides a highly sensitive and specific tool to support GPs in diagnosing depression. Future algorithm improvements may integrate further biological markers and longitudinal data. The tool's clinical utility needs further evaluation through a randomized controlled trial, which is currently being planned. Additionally, assessing whether GPs actively integrate the algorithm's predictions into their diagnostic and treatment decisions will be critical for its practical adoption.
AD - Department of Psychiatry and Psychotherapy, LMU University Hospital, LMU Munich, Nussbaumstraße 7, 80336, Munich, Germany. j.eder@med.uni-muenchen.de.; Graduate Program "POKAL - Predictors and Outcomes in Primary Care" (DFG-GrK 2621), Munich, Germany. j.eder@med.uni-muenchen.de.; German Center for Mental Health (DZPG), Partner Site Munich-Augsburg, Munich, Germany. j.eder@med.uni-muenchen.de.; Department of Psychiatry and Psychotherapy, LMU University Hospital, LMU Munich, Nussbaumstraße 7, 80336, Munich, Germany.; Graduate Program "POKAL - Predictors and Outcomes in Primary Care" (DFG-GrK 2621), Munich, Germany.; German Center for Mental Health (DZPG), Partner Site Munich-Augsburg, Munich, Germany.; Chair for Information Systems, Technical University of Munich (TUM), 85748, Garching, Germany.; Institute of Psychiatry, Psychology and Neuroscience, King's College London, London, UK.; Institute of General Practice and Health Services Research, School of Medicine, Technical University Munich, Munich, Germany.; Institute of General Practice and Family Medicine, Ludwig-Maximilians-University Munich, Munich, Germany.; Oberberg Fachklinik Bad Tölz, Bad Tölz, Germany.; Max-Planck Institute of Psychiatry, Munich, Germany.
AN - 40063259
BT - Eur Arch Psychiatry Clin Neurosci
C5 - HIT & Telehealth
CP - 2
DA - Mar
DO - 10.1007/s00406-025-01990-5
DP - NLM
ET - 20250310
IS - 2
JF - Eur Arch Psychiatry Clin Neurosci
LA - eng
N2 - General practitioners play an essential role in identifying depression and are often the first point of contact for patients. Current diagnostic tools, such as the Patient Health Questionnaire-9, provide initial screening but might lead to false positives. To address this, we developed a two-step machine learning model called Clinical 15, trained on a cohort of 581 participants using a nested cross-validation framework. The model integrates self-reported data from validated questionnaires within a study sample of patients presenting to general practitioners. Clinical 15 demonstrated a balanced accuracy of 88.2% and incorporates a traffic light system: green for healthy, red for depression, and yellow for uncertain cases. Gaussian mixture model clustering identified four depression subtypes, including an Immuno-Metabolic cluster characterized by obesity, low-grade inflammation, autonomic nervous system dysregulation, and reduced physical activity. The Clinical 15 algorithm identified all patients within the immuno-metabolic cluster as depressed, although 22.2% (30.8% across the whole dataset) were categorized as uncertain, leading to a yellow traffic light. The biological characterization of patients and monitoring of their clinical course may be used for differential risk stratification in the future. In conclusion, the Clinical 15 model provides a highly sensitive and specific tool to support GPs in diagnosing depression. Future algorithm improvements may integrate further biological markers and longitudinal data. The tool's clinical utility needs further evaluation through a randomized controlled trial, which is currently being planned. Additionally, assessing whether GPs actively integrate the algorithm's predictions into their diagnostic and treatment decisions will be critical for its practical adoption.
PY - 2026
SN - 0940-1334 (Print); 0940-1334
SP - 407
EP - 420+
ST - A multimodal approach to depression diagnosis: insights from machine learning algorithm development in primary care
T1 - A multimodal approach to depression diagnosis: insights from machine learning algorithm development in primary care
T2 - Eur Arch Psychiatry Clin Neurosci
TI - A multimodal approach to depression diagnosis: insights from machine learning algorithm development in primary care
U1 - HIT & Telehealth
U3 - 10.1007/s00406-025-01990-5
VL - 276
VO - 0940-1334 (Print); 0940-1334
Y1 - 2026
ER -