Predicting Self-harm Incidents Following Inpatient Visits using Disease Comorbidity Network

Self-harm is serious but preventable, particularly if the risk can be identified early. But the early detection of self-harm individuals is so far not satisfied. This study aims to develop and test a comorbidity network-enhanced deep learning framework to improve the prediction of individual self-harm within 12 months after hospital discharge. Between January 1, 2007, and December 31, 2010, we obtained 2,323 patients with self-harm clinical record from 1,764,094 inpatients across 44 public hospitals in Hong Kong and 46,460 randomly sampled population controls from those same hospitals. Eighty percent (80%) of the study sample was randomly selected for model training, and the remaining 20% was set aside for model testing. The proposed comorbidity network-enhanced model was compared with a baseline deep learning model for self-harm prediction. The C-statistic, precision and sensitivity were used to evaluate the prediction accuracy of the proposed model and the baseline model. Experiments demonstrated that the proposed comorbidity network-enhanced model outperformed baseline model in identifying patients who would self-harm within 12 months (C-statistic of proposed model 0.89). The precision was 0.54 for positive cases and 0.98 for negative cases, whilst the sensitivity was 0.72 for positive cases and 0.96 for negative cases. Results indicated that it is critical to consider the general disease comorbidity patterns in self-harm screening and prevention programs. The model also extracted the most predictive diagnoses, and pairs of comorbid diagnoses which provide medical professionals with an effective screening strategy.