Predicting SILENT after lithium poisoning: a multicenter cohort study and pragmatic risk stratification model

Abstract

INTRODUCTION: The Syndrome of Irreversible Lithium-Effectuated Neurotoxicity (SILENT) is a rare but potentially devastating complication of lithium poisoning. Current evidence rests on case reports and small series, leaving its true incidence, risk factors, and follow-up tools poorly defined. METHODS: Retrospective, international, multicenter cohort of adults evaluated for lithium poisoning (2015-2022) across 23 tertiary hospitals in Spain, Ireland, and Switzerland. SILENT was defined as persistent neurological sequelae attributed to lithium toxicity documented at 60 days. A parsimonious multivariable logistic regression model was built from routinely available clinical variables and assessed by ROC analysis and internal split-sample validation. RESULTS: Among 644 episodes, SILENT was identified in 49 (7.6%); predicted probabilities were available for 456 episodes (39 SILENT cases). Affected patients were older and more often had acute-on-chronic poisoning, early neurological manifestations, and greater clinical severity. Serum lithium concentrations were not associated with SILENT. The model-age, poisoning type, neurological findings (ataxia/myoclonus), Poisoning Severity Score, and peak urea-showed good discrimination (AUC 0.813; 95% CI 0.743-0.882) and acceptable internal validation (AUC 0.730; 95% CI 0.572-0.888). A sensitivity-oriented threshold (P≥0.0475) yielded a high negative predictive value (98.2%), supporting its use to prioritize post-discharge neurological follow-up. DISCUSSION: SILENT occurred more frequently than case-based literature suggests and followed a recognizable presentation pattern. Serum lithium alone did not predict long-term sequelae. The model, based on routine variables, may offer a pragmatic tool for early risk stratification. CONCLUSION: SILENT is not an exceptional outcome after lithium poisoning and cannot be predicted by serum lithium alone. A simple clinical prediction model may identify patients warranting structured neurological follow-up. External validation is required before implementation.