A machine learning approach to risk assessment for alcohol withdrawal syndrome

At present, risk assessment for alcohol withdrawal syndrome relies on clinical judgment. Our aim was to develop accurate machine learning tools to predict alcohol withdrawal outcomes at the individual subject level using information easily attainable at patients' admission. An observational machine learning analysis using nested cross-validation and out-of-sample validation was applied to alcohol-dependent patients at two major detoxification wards (LMU, n = 389; TU, n = 805). 121 retrospectively derived clinical, blood-derived, and sociodemographic measures were used to predict 1) moderate to severe withdrawal defined by the alcohol withdrawal scale, 2) delirium tremens, and 3) withdrawal seizures. Mild and more severe withdrawal cases could be separated with significant, although highly variable accuracy in both samples (LMU, balanced accuracy [BAC] = 69.4%; TU, BAC = 55.9%). Poor outcome predictions were associated with higher cumulative clomethiazole doses during the withdrawal course. Delirium tremens was predicted in the TU cohort with BAC of 75%. No significant model predicting withdrawal seizures could be found. Our models were unique to each treatment site and thus did not generalize. For both treatment sites and withdrawal outcome different variable sets informed our models' decisions. Besides previously described variables (most notably, thrombocytopenia), we identified new predictors (history of blood pressure abnormalities, urine screening for benzodiazepines and educational attainment). In conclusion, machine learning approaches may facilitate generalizable, individualized predictions for alcohol withdrawal severity. Since predictive patterns highly vary for different outcomes of withdrawal severity and across treatment sites, prediction tools should not be recommended for clinical practice unless adequately validated in specific cohorts.