CloudMedx Logo

CloudMedx Predicts Survival in ALS with 82% accuracy based on clinical markers – Published by Barrow Neurological Institute

Bill JacobsenShafeeq Ladha

Title: Predicting Survival in ALS Using Machine Learning Algorithms – a Preliminary Analysis (5374)


Objective: Use machine learning to better predict survival outcomes in amyotrophic lateral sclerosis (ALS).

Background: ALS is a progressive degenerative motor neuron disease. Due to the heterogeneity of the disease and difficulty predicting disease progression, interventional trials are lengthy and require large sample sizes to account for this. Tools that better stratify patients based on predicted survival or rate of disease progression would increase our ability to design shorter and cheaper clinical trials with a higher chance of success.

Design/Methods: In collaboration with CloudMedx Inc we analyzed the publicly available Pooled Resource Open-Access ALS Clinical Trials (PROACT) database to develop a machine learning framework to identify variables with high predictive power for survival stratification. Seven hundred twenty-six patients with complete records were divided into three buckets for survival and normalized to 242 patients per bucket. Running algorithms of k-nearest neighbors (kNN), support vector machine (SVM), random forest model, multilayer perceptron (MLP), and gradient boosting machine (GBM), this model was trained, validated, and tested with initial visit data and 90-day data to develop an outcome prediction algorithm.

Results: The CloudMedx ALS Framework (CMX-ALS) identified twenty highly predictive variables for survival. Data from 90-day and initial visit predicted survival outcomes with 82% and 70% accuracy respectively. This stratified patients into survival groups of less than 1 year, 1–2 years, and greater than 2 years. We identified seven new lab variables (bicarbonate, hematocrit, potassium, chloride, AST, glucose, and hemoglobin) and two medications (amitriptyline and baclofen) that were highly predictive of survival outcomes, and have not previously been mentioned in survival prediction models.

Conclusions: Patient stratification using machine learning methods will allow for smaller sample sizes in trial design. Additionally, we identified new variables and medications that were not previously considered for outcome efficacy.

For complete white paper, please visit: