PsychoPharm aggregated risk score (PARS): a multidimensional tool to flag high-risk pharmacotherapy in psychiatric...

Psychiatric outpatients frequently receive complex, long-term regimens where polypharmacy, drug-drug interactions (DDIs), QT interval risk, anticholinergic load, and serotonergic exposure co-occur. Clinicians must triage drug lists quickly, yet most tools address a single risk domain. We developed the PsychoPharm Aggregated Risk Score (PARS), a visit-level composite score that integrates multiple pharmacological risk domains into a single, clinically interpretable signal.

We analyzed 2,666 visits from 680 adults in ambulatory psychiatry. For each visit, we computed the risk components: DDI score from DrugBank, AZCERT QT risk, anticholinergic cognitive burden (ACB), serotonergic exposure, and polypharmacy ( $\ge$ 5 drugs). Using $z$ -standardized components, we created the PARS composite as their equal-weight mean, and report it as PARS rescaled on a 0-10 scale for clinical communication. We trained a logistic regression to estimate the probability of High-risk PARS, defined as PARS $\ge$ 75th percentile, using only prescribed drugs, age, and sex as features (explicitly excluding the engineered components to prevent leakage), then compared with additional stratifications by age and sex, as well as a 5-fold GroupKFold. A fixed decision threshold of 0.30 was set to favor recall and F1 for the positive class. A drug-only model was also used to identify drugs associated with High-risk visits.

The prevalence of high-risk cases was observed in 25.2% of visits. Across validation schemes, the primary 80-20 patient-level split achieved the best accuracy, precision, and F1 at the 0.30 operating point. In the primary 80-20 split stratified by High-risk status, the logistic model achieved an AUC of 0.931 (patient-bootstrap 95% CI 0.875-0.971). Discrimination was similar with additional age and sex stratification (AUC of 0.938) and 5-fold GroupKFold (pooled out-of-fold AUC 0.939), indicating robustness to partitioning. In the drug-only model, positive associations included drugs such as quetiapine, haloperidol, clozapine, and amiodarone. The highest-ranked visits combined central nervous system-heavy regimens and polypharmacy.

As an exploratory tool, PARS integrates DDIs, QT risk, anticholinergic cognitive burden, serotonergic exposure, and polypharmacy into a single probability that reliably discriminates High-risk visits and supports screening at a 0.30 operating threshold. Our approach highlights actionable drug combinations and patient profiles for drug review and deprescribing.