Development of a machine learning-based survival prediction model for ALS inclusive of the advanced-stage population

OBJECTIVE: ]). METHODS: = 403). Baseline patient characteristics, including slopes from symptom onset, were used to train a random forest model to identify parameters with the greatest relative importance (RI) for predicting survival outcomes. These parameters were used to train a gradient-boosting machine (GBM) model that generated patient-level survival predictions (log-hazard). Model discrimination and calibration were quantified by C-index and calibration-in-the-large plus calibration slope, respectively. Kaplan-Meier curves were generated, with patient stratification into tertiles based on the predicted survival risk score. RESULTS: Baseline characteristics with the highest RI for driving survival predictions included: VC% slope (20.2%); age (12.4%); VC% (9.9%); VC(L) (7.5%); ALSFRS-R (6.6%); and ALSFRS-R slope (5.1%). Model performance upon external validation was satisfactory for both discrimination (C-index, 0.709 [95% CI, 0.671-0.746]) and calibration (calibration-in-the-large, 0.083 [95% CI, -0.073-0.232]; calibration slope, 0.992 [95% CI, 0.789-1.198]). At 8-months from baseline, the model successfully stratified patients by survival prognosis, with low-, average-, and high-risk population tertiles having observed median survival probabilities of 85, 69, and 43%, respectively. CONCLUSIONS: This model accurately predicts survival prognosis in ALS, including patients with severely impaired respiratory function. This new understanding of patient-specific factors that drive survival prognostication will be invaluable for reducing patient heterogeneity in clinical trials evaluating novel therapeutic modalities in early- and advanced-stage ALS.