Post-Traumatic Stress Disorder (PTSD) is one of the most common mental health disorders prevalent in the US. Most alarming, PTSD occurs at double the rate for combat veterans compared to the general population. Severity of PTSD is associated with risk taking behaviors such as substance abuse, non-suicidal self-injury, sexual risk behaviors, among other negative behaviors. Psychological disorders are often preceded by crisis events, thus monitoring for crisis events can help prevent risky behavior in veterans. Ecological momentary assessment techniques are effective in capturing possible crisis events for veterans. Mobile apps are commonly used to gather such behavioral changes in participants. Crisis events collected from m-health can be analyzed for the identification of long- term PTSD risk. Early identification of risk can help in planning intervention to mitigate the risk. Many scholars have used traditional statistical and machine learning methods for the prediction of mental health issues in individuals. But these models lack transparency in how decisions are made. Providing justifications for the predictions can increase the reliability of the model. Our research focused on developing an explainable prediction model using class association rules to identify veterans at risk of persistent PTSD. The generated association rules serve as precursors to the long-term crisis in veterans. Results of the analysis showed that having no family support, little or no interest in hobbies, stress and lack of sleep are some of the influencing factors of persistent PTSD in veterans.

People are prone to develop opioid dependence and other health problems due to regular non-medical use, prolonged use, and misuse of opioids. The number of hospital admissions for opioid dependence is growing across the US. The length of stay (LOS) is an essential indicator that assesses the severity of opioid overdose admissions. In this paper, opioid-related healthcare data from Froedtert Health Medical System in Wisconsin are analyzed and machine learning models are proposed to predict the LOS of opioid overdose admitted patients. We also determine important features that impact the LOS. To explore the factors that significantly influence the LOS, we implemented recursive feature elimination (RFE) to select important features from the data. Since the data set is imbalanced, we applied two imbalanced learning approaches to tackle that, namely, SMOTE and imbalanced learning models. Several machine learning models were constructed and validated with 10 iterations of 10-fold cross validation, and we fine-tuned the models with the highest f1 score and AUC score. Random Forest Classifier outperforms other models trained on the oversampled data set (AUC = 0.81, precision = 0.72, recall = 0.69). Easy Ensemble Classifier, which is trained on the imbalanced data set, has a better performance in predicting longer stays (AUC = 0.80, precision = 0.70, recall = 0.74). Finally, important features identified from the predictive models can be used to determine at-risk patients of longer LOS and provide appropriate preventive care and resources.

Opioid use disorder (OUD) is a medical condition associated with problematic patterns of opioid use that cause interpersonal and social impairment. This research demonstrates how supervised machine learning can be used to predict patients at risk of hospital readmission following opioid overdose, and to predict patients at risk of developing OUD. Two labeled datasets were built from deidentified hospital data provided by a Level I Trauma Center Hospital. Several machine learning models were constructed (logistic regression, random forest, support vector machine, AdaBoost, XGBoost) and validated with 10 iterations of 10-fold cross validation. The XGBoost classifier can sufficiently predict patients at risk for OUD (AUC = 0.78, precision = 0.71, recall = 0.53). This work can assist providers in determining appropriate preventive care and resources for at-risk patients.