Self-HArm one-year Risk Prediction in personality disorders (SHARP-PD): a clinical risk prediction model development and internal-external cross-validation study

Background.Personality-disorders (PDs) are associated with markedly-increased risk of selfharm.We aimed to develop and validate a risk-prediction model (Self-HArm one-year Risk Prediction in personality disorders [SHARP-PD]) to predict 1-year self-harm in people with newly-diagnosed PD.Methods.This population-based study identified individuals with first-recorded diagnosis of PD in 2002-2021, utilising an electronic-health-record database of public-healthcare service in Hong-Kong.SHARP-PD was developed using binary-logistic-regression model, with Least-Absolute-Shrinkage-and-Selection-Operator penalisation.Model performance was assessed through internal-external cross-validation (IECV), sequentially leaving out one cluster from the seven healthcare-catchment areas for validation and averaging performance across seven clusters via random-effect meta-analysis.Predictors for model development included socio-demographic characteristics, psychiatric and physical morbidities, psychotropic medication use, history of selfpoisoning and self-inflicting injuries, and psychiatric service utilisation.Results.Among 6229 patients with PD, 608 (9.8%) performed self-harm at 1-year post-diagnosis.The model exhibited excellent performance in IECV: area-under-the-curve 0.93 (0.92-0.94);Brier-score 0.06 (0.05-0.06); calibration-slope 1.01 (0.95-1.08); calibration-intercept 0.02 (-0.14-0.19);integrated-calibration-index 0.15 (0.11-0.19); observed/expected-ratio 1.00 (0.92-1.08).Decision-curve-analysis showed that SHARP-PD provided greater net-benefit than the treat-all and treat-none strategies starting from the threshold probability of 1%.The model performance remained largely-consistent in sensitivity-analyses excluding predictors with near-zero prevalence, complete-case analyses, including longer washout period for incident PD diagnosis, and restricting to non-fatal self-harm in outcome definition, and across key socio-demographic subgroups of sex, age categories, and borderline PD.