結合邏輯斯迴歸與機器學習探討職傷後三個月的心理健康對十二個月未復工之預測應用

Abstract

研究目的：本研究旨在探討受到職業傷害的勞工在三個月後的心理健康狀態是否能有效預測十二個月時無法復工的情形。並藉由傳統邏輯斯迴歸及機器學習方法，進一步分析各心理情況指標的預測力，期望能應用於高風險個案的早期發現與介入策略。 方法：本研究納入2009年在臺灣因職業傷害住院三天以上並領取職災住院給付之勞工，並追蹤3個月和12個月的資料調查，並以簡式健康量表（BSRS-50）評估心理健康狀態。數據分析採用邏輯斯迴歸分析及機器學習方法（決策樹、隨機森林、XGBoost）等模型，透過上採樣和成本敏感分析，以瞭解心理情況與未復工的相關性因素。 結果：在單變項分析中，人口學資料、心理健康情況與未復工的結果呈現顯著相關，然而在多變項模型中，僅部分人口學資料仍呈現顯著影響復工的情形。在機器學習上採樣的模型策略下，隨機森林表顯最為平衡，而在成本敏感學習下，XGboost與隨機森林模型對於未復工的個案敏感度較佳。而後續進行SHAP分析中亦顯示部分心理症狀（如恐懼、附加症狀、精神病性、憂鬱等）會對預測結果產生影響。 結論：心理健康問題在職業傷害後早期發現能對復工的結果呈現相關，雖然傳統統計分析中在控制其他變項效應不顯著，但在機器學習模型中仍有預測價值。顯著，但在機器學習模型中仍有預測價值。透過模型的預測分析，有助於未來應用於個案早期發展個人化的介入策略，能提供個案更好的介入資源。

Objective: This study aimed to examine whether the psychological status of workers three months after an occupational injury could effectively predict non-return-to-work outcomes at twelve months. Using both traditional logistic regression and machine learning approaches, with the goal of informing early identification and intervention strategies for high-risk cases. Methods: The study included workers in Taiwan who were hospitalized for more than three days due to occupational injuries and received Inpatient Hospitalization Benefit of Occupational Accident Medical Benefits under the Labor Insurance in 2009. Data were collected at three and twelve months after injury. Psychological status was assessed using the Brief Symptom Rating Scale (BSRS-50). Data analysis employed logistic regression as well as machine learning models, including decision tree, random forest, and XGBoost. Up-sampling and cost-sensitive learning techniques were applied to address data imbalance and to explore associations between psychological factors and non-return-to-work outcomes. Results: In single logistic analysis, demographic and psychological variables were significantly associated with non-return-to-work outcomes. However, in the multiple model, only a subset of demographic factors remained statistically significant predictors. Among the up-sampling models, the random forest model showed the most balanced performance. Under the cost-sensitive learning approach, XGBoost and random forest models showed higher sensitivity in identifying non-return-to-work cases. Moreover, SHAP analysis also revealed that certain psychological symptoms, such as phobic anxiety, additional symptoms, psychoticism, and depression, contributed meaningfully to the prediction outcomes. Conclusion: Early identification of psychological health symptoms after occupational injury is associated with return-to-work outcomes. Although traditional logistic model did not show significant effects for psychological variables after adjusting for covariates, machine learning models retained predictive value. These predictive models may facilitate the development of personalized early interventions and improve access to appropriate support for injured workers.