運用機器學習預測方式輔助醫師瞭解精神疾病犯罪者監護處分後是否再犯以提供決策支援

Machine Learning-Based Recidivism Prediction for Offenders with Mental Illness Following Custodial

臺灣雖有全民健保料庫，但如果牽涉司法案件的醫療處置，還需要司法端的資訊才能知道醫療結束後病患的確實狀況，因缺乏醫學法律相關資料庫，所以如本研究追蹤74位患者之後續變化的相關文獻極為少見，為此本研究借助利用機器學習的方法，分析與預測監護處分後精神疾病犯罪者是否再犯以對其領域做出貢獻。相關資料來源為某精神專科教學醫院司法精神醫學資料庫，運用2010至2020年中長期追蹤研究74位有監護處分的精神疾病犯罪者1到11年資料，並將資料切分為80%作為訓練資料，剩餘20％資料作為驗證資料，本研究在為減少因資料量少且類別分布不平均的情況所引發的疑慮，於研究的實作中有進行SMOTE+ENN方法，並運用機器學習方法訓練模型。最後在運用驗證資料驗證機器學習方法建立相關再犯率的預測模型，並使用F1-score作為模型效能評估方式，由模型F1-score的評估方式的結果中得知（Extremely randomized trees，ExtraTrees）Classifier為0.86是所有機器學習模型中最好的，以此模型做為我們預測精神疾病犯罪者監護處分後再犯情形的預測模型，即可以達成最佳的預測結果有效對輔助降低再犯率做出貢獻，藉以達到輔助醫護人員做出最佳醫療決策，以此提供早期預測降低再犯率情形可提升精神疾病該領域醫療品質。

Although Taiwan has a comprehensive National Health Insurance database, medical dispositions involving judicial cases require additional information from the judicial system to determine the actual status of patients after treatment. Due to the lack of integrated medico-legal databases, studies tracking post-treatment changes among offenders with mental illness such as the present research involving 74 patients are extremely rare. To address this gap, this study employs machine learning techniques to analyze and predict whether offenders with mental illness will reoffend following custodial treatment, thereby contributing to the field.
Data were obtained from the forensic psychiatry database of a specialized teaching hospital. A longitudinal dataset covering 74 offenders with mental illness subjected to custodial measures between 2010 and 2020 was used, with follow-up periods ranging from 1 to 11 years. The dataset was split into 80% for training and 20% for validation. To mitigate concerns arising from limited sample size and class imbalance, the SMOTE+ENN method was applied during implementation, followed by machine learning model training.
Finally, validation data were used to evaluate predictive models for recidivism using F1-score as the performance metric. Among all tested models, the Extremely Randomized Trees (ExtraTrees) Classifier achieved the highest F1-score of 0.86. This model was therefore selected as the optimal predictive tool for assessing recidivism risk among offenders with mental illness after custodial treatment. By enabling early prediction and supporting clinical decision making, this approach has the potential to reduce recidivism rates and improve the quality of psychiatric care in this domain.