Development of an explainable machine learning model for predicting six-month consciousness recovery in patients with prolonged disorders of consciousness

Abstract

With advances in emergency care systems, the number of patients with prolonged disorders of consciousness (pDoC) has markedly increased. To optimize the allocation of healthcare resources and mitigate risks, the early and accurate identification of recovery potential in pDoC is essential. Current machine learning models are often hindered by small sample sizes, inconsistent data standards, and limited interpretability. This study aims to overcome these issues and develop a machine learning-based prognostic model that may provide potential support for clinical decision-making in patients with pDoC. A retrospective study of pDoC patients hospitalized at Xuanwu Hospital from January 2013 to August 2024 was conducted. Data on clinical information, behavioral scales, and event-related potentials (ERPs) were collected, using 6-month consciousness recovery as the outcome. Two auditory oddball paradigms with progressively increasing difficulty were used to elicit ERPs, in which the subject’s own name (SON) was presented as the deviant stimulus, and a 1000 Hz pure tone together with the subject’s derived name (SDN) served as the standard stimuli respectively. The least absolute shrinkage and selection operator (LASSO) regression analysis selected eight variables: sex, skull state, etiology, duration, Coma Recovery Scale-Revised (CRS-R) score, diagnosis, mismatch negativity (MMN) pattern, and epilepsy status. Model performance was assessed using metrics such as macro-averaged receiver operating characteristic area under the curve (ROC-AUC), macro-averaged precision–recall area under the curve (PR-AUC) and macro-averaged F1 score. Clinical utility was evaluated via Decision Curve Analysis (DCA), and interpretability was supported by SHapley Additive exPlanations (SHAP) method. A total of 110 patients with pDoC, comprising 72 with minimally conscious state (MCS) and 38 with vegetative state/unresponsive wakefulness syndrome (VS/UWS), were enrolled in the study. Follow-up assessments identified 53 patients with improved prognosis and 57 without improvement. The efficient logistic regression model showed moderate discriminative ability with a macro-averaged ROC-AUC of 0.7504 (95% CI: 0.6590–0.8404). The CRS-R score ranked highest in SHAP importance. In the decision curve analysis, the threshold probabilities for net benefit ranged from 13 to 89%. This study developed a machine learning-based prognostic model for consciousness recovery in pDoC patients by integrating multimodal bedside data using efficient logistic regression. The model achieved moderate discriminative performance and provides preliminary evidence supporting individualized outcome predictions. When combined with clinical judgment, it may assist in optimizing treatment decisions for pDoC patients, although further validation is required.