Abstract Details

Title Predicting Fast Progressors in Amyotrophic Lateral Sclerosis (ALS): A Machine Learning Approach

Topic Neuromuscular and Clinical Neurophysiology (EMG)

Presentation(s) P1 - Poster Session 1 (9:00 AM-5:00 PM)

Poster/Presentation
Number 389

Objective Develop machine learning algorithms to predict ALS patients who are fast progressors and address challenges with time-series forecasting, bias and missing data inherent in clinical databases.

Background Variability in ALS progression poses challenges in detecting decline and treatment response. Predicting fast disease progressors can enrich clinical trials with patients most likely to demonstrate treatment effect over a restricted time-period. Based on recent edaravone and AMX0035 trials, we defined such patients as declining more than 1.5 points/month on the revised ALS functional rating scale (ALSFRS-R).

Design/Methods Pooled clinical, demographic and lab data of 4328 ALS participants from the PRO-ACT database with ≥50% available data points and ≥2 visits during the preceding 3, 6, and 9 months were used to predict ALSFRS-R progression over the subsequent 3, 6, and 9 months. Participants were classified as fast progressors (ALSFRS-R > –1.5/month) or non-fast progressors, using an ensemble of XGBoost and Bayesian Long Short-Term Memory (BLSTM) models, with a 80:20 train:test split. BLSTM was used to model temporal dependencies and provide the confidence of prediction. Missing data were imputed using polynomial regression and k-nearest neighbors in the XGBoost model, and simultaneous imputation and prediction in BLSTM model.

Results XGBoost predicted fast progressors with an accuracy of 0.63-0.74 across the 9 models. BLSTM had an accuracy of 0.72 and was very confident in 87% of its accurate predictions. The model ensemble had an accuracy of 0.71. The top predictive parameters, which appeared at least 3 times across all models included well-known clinical prognostic features such as past ALSFRS-R, FVC and time since symptom onset. It also included non-standard predictive features such as blood AST, calcium and creatinine for further study.

Conclusions Machine learning algorithms accurately predicted fast progressors, identified predictive features, and addressed missing values and variable follow-up in large clinical databases.

Authors/Disclosures
PRESENTER	No disclosure on file
	No disclosure on file
Erik P. Pioro, MD, DPhil, FAAN (University of British Columbia)	Dr. Pioro has received personal compensation in the range of $500-$4,999 for serving as a Consultant for Avanir Pharmaceutical, Inc.. Dr. Pioro has received personal compensation in the range of $500-$4,999 for serving as a Consultant for Amylyx Pharmaceuticals. Dr. Pioro has received personal compensation in the range of $500-$4,999 for serving as a Consultant for Argenx. Dr. Pioro has received personal compensation in the range of $500-$4,999 for serving as a Consultant for MT Pharma America, Inc.. Dr. Pioro has received personal compensation in the range of $500-$4,999 for serving on a Scientific Advisory or Data Safety Monitoring board for NeuroTherapia, Inc.. Dr. Pioro has received personal compensation in the range of $500-$4,999 for serving on a Scientific Advisory or Data Safety Monitoring board for MT Pharma America, Inc..
	No disclosure on file
Crystal J. Yeo, MD, PhD	Dr. Yeo has nothing to disclose.

��ɫ��

��ɫ��