Background: Coercive measures such as seclusion or mechanical restraint remain in routine use in German psychiatric hospitals despite ethical controversy. Reliable, data-driven risk stratification could help clinicians prevent their use. Objective: To identify and predict risk factors for coercive measures in acute inpatient psychiatry by combining electronic health records (EHRs) with large language models (LLMs) and classical machine learning. Methods: We retrospectively analysed 2,320 admissions (2022-2024) to the psychiatric emergency ward at a German university clinic psychiatry department. Free-text discharge summaries were processed with Llama-3–based chat models. Through prompt-engineering and in-context learning, the models automatically extracted five binary or categorical predictors—psychosis, suicidality, aggression, police involvement, and intoxication. Extraction performance was first validated on 100 randomly selected cases without coercion. A linear support-vector machine (SVM) was then trained to classify episodes with vs. without coercive measures, using stratified 5-fold cross-validation and minority oversampling. All procedures complied with the Declaration of Helsinki and were approved by the TU Dresden ethics committee. Results: LLM extraction accuracies in the validation set were 92 % (police involvement), 90 % (suicidality), 87 % (intoxication), 85% (psychosis), and 82% (aggression), respectively. Of the 2,320 cases, 193 (8.3 %) involved coercion. The SVM achieved a mean balanced accuracy of 71 %. Police involvement (79 %), aggression (77 %), psychosis (74 %), and suicidality (63 %) were the strongest individual predictors. Conclusions: Automated LLM-based feature extraction from routine EHR text, coupled with an SVM classifier, provides a feasible and accurate approach to forecasting coercive measures. Integrating such models into clinical workflows could enable real-time identification of high-risk patients and support targeted de-escalation strategies.