Machine Learning Based Models To Predict Functional Improvement in Stroke Patients During Inpatient Rehabilitation
| Autor: | Sara Ali, Dorothea Parker, Shayan Shams, Yan Chu, Syed Zamin, Xiaoqian Jiang, Sean I Savitz, Lisa W. Thomas, Joseph Wozny |
|---|---|
| Rok vydání: | 2021 |
| Předmět: |
Intracerebral hemorrhage
Rehabilitation Receiver operating characteristic business.industry medicine.medical_treatment Psychological intervention Physical Therapy Sports Therapy and Rehabilitation Retrospective cohort study medicine.disease Machine learning computer.software_genre Functional Independence Measure Case mix index Medicine Artificial intelligence business Stroke computer |
| Zdroj: | Archives of Physical Medicine and Rehabilitation. 102:e56 |
| ISSN: | 0003-9993 |
| DOI: | 10.1016/j.apmr.2021.07.633 |
| Popis: | Research Objectives To build a predictive model and find demographic and clinical factors associated with functional improvement in acute stroke survivors using Natural Language Processing (NLP) and Machine Learning (ML). Design Retrospective study of clinical and demographic factors and outcomes. Setting Acute stroke patients admitted to Memorial Hermann/TIRR inpatient rehabilitation facilities. Participants A convenience sample that includes individuals admitted from 2017-2019 and excludes individuals with multiple stroke or pre-stroke neurological conditions. It documents 421 male and 382 female with median age of 69 and racial breakdown of 43% Caucasian, 28% African American, and 18% Asian. Right and left laterality are equal. Bilateral stroke noted in 73 clients. One-quarter of patients had intracerebral hemorrhage and the rest had ischemic stroke. Interventions N/A. Main Outcome Measures We developed a novel model using gradient-boosting tree algorithm, CatBoost, and utilized Shapley values to estimate the contribution of each predictor to significant functional improvement (SFI). SFI was defined as an increase by a minimum of two points on at least 5 FIM (Functional Independence Measure) subset scores from admission to discharge. We analyzed 229 Independent variables including demographics, stroke characteristics, severity on the NIHSS, Initial FIM scores at admission, NLP vectorized clinical notes, history of hypertension, smoking, diabetes, and hyperlipidemia. Results Our model achieved an area under receiver operator curve of 0.89 showing better performance compared to commonly used ML and statistical models. Based on Shapely values, age, presenting symptoms, stroke laterality, initial NIHSS, case mix index score and inpatient therapy time have a more pronounced role in predicting functional improvement. Conclusions ML models carry the high potential to analyze large multidimensional data on stroke patients undergoing inpatient rehabilitation. They may help clinicians better predict functional improvement for patients undergoing rehabilitation. Author(s) Disclosures The authors declare no conflict of interest. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|