Predictive Performance of Machine Learning–Based Models for Poststroke Clinical Outcomes in Comparison With Conventional Prognostic Scores: Multicenter, Hospital-Based Observational Study (Preprint)

• DOI: 10.2196/preprints.46840
• Publication Date: 2023-03-09T19:14:53Z
• Authors: Fumi Irie; Koutarou Matsumoto; Ryu Matsuo; Yasunobu Nohara; Yoshinobu Wakisaka; Tetsuro Ago; Naoki Nakashima; Takanari Kitazono; Masahiro Kamouchi
• Keywords: Stroke

<sec> <title>BACKGROUND</title> Although machine learning is a promising tool for prognostication, the performance of in predicting outcomes after stroke remains to be examined. </sec> <title>OBJECTIVE</title> We aimed examine how much data-driven models with improve predictive post-stroke compared conventional prognostic scores and elucidate explanatory variables learning-based differ from items scores. <title>METHODS</title> used data 10513 patients registered multicenter prospective registry Japan between 2007 2017. The were poor functional outcome (modified Rankin Scale score&gt;2) death at 3 months post-stroke. Machine developed using all regularization methods, random forests, or boosted trees. selected three scores, namely ASTRAL (Acute STroke Registry Analysis Lausanne) score, PLAN (Preadmission comorbidities, Level consciousness, Age, Neurologic deficit) iScore comparison. Item-based regression these Model was assessed terms discrimination calibration. To compare model that item-based model, we performed internal validation splits identical populations into 80% as training set 20% test set: validated set. evaluated contribution each variable predictors <title>RESULTS</title> mean (SD) age study 73.0 (12.5) years, 59.1% them men. area under receiver operating characteristic curves precision-recall higher than splits. also better Brier score. different variables, such laboratory data, Including improved stroke, especially death. <title>CONCLUSIONS</title> though they required additional attain performance. Further studies are warranted validate usefulness clinical settings.