By Frank Fuller

‍

We are excited to announce the latest release of our machine-learning model for Amyotrophic Lateral Sclerosis (ALS), a progressive neurodegenerative disease that affects nerve cells in the brain and spinal cord. 

Our new model, ALS DTG 3.0, is designed to accelerate ALS clinical trials by providing a generative model of each patient's future clinical trial outcomes. By helping researchers differentiate treatment effects from noise more clearly, this model reduces both the costs and timelines associated with clinical trials.

‍

What's New in This Release

This latest release leverages a couple of new additions to the Neural Boltzmann Machine architecture:

• Multivariate Event Modeling: Our previous model could only forecast a single event—death. This release adds support for four critical events: death, respiratory failure, gastrostomy, and tracheostomy. Monitoring these events allows researchers to assess patient progression earlier, ultimately speeding up trial timelines.
  ‍
• Improved Event-Longitudinal Interdependence: In contrast to the earlier versions, where the longitudinal outcomes and event models were independent, the new model incorporates conditional dependencies between these outcomes. This improvement enhances the Combined Assessment of Function and Survival (CAFS), a metric that leverages both ALS Functional Rating Scale (ALSFRS) scores and death events. CAFS is increasingly used in modern ALS trials to provide a clearer view of how a patient is progressing relative to their peers.

In addition, we have expanded the training data set:

• Expanded Data Coverage: The model now incorporates data from the Answer-ALS observational trial, netting an additional 250 patients. This inclusion helps the model represent the ALS population more comprehensively.
  ‍
• Enhanced Feature Set: We’ve expanded the range of clinical features to include the ALS Cognitive Behavioral Screen, Body Mass Index (BMI), and (ALSFRS) respiratory scores. These additions improve the model’s ability to forecast clinical outcomes by capturing a more complete picture of patient health.

Performance Improvements

Our new model demonstrates significant improvements in performance, especially in discriminative tasks:

‍Higher Discriminative Performance: We observed a marked improvement in correlation metrics on validation data, leading to better “discriminative” performance in both the validation and test sets compared to the previous release.

C-Index Gains on Death Prediction: The model shows notable gains in the C-index for predicting death, indicating its ability to accurately rank patients by likelihood of mortality. However, we observed a slight regression in the precision of predicted survival probabilities, suggesting that while the model excels at ranking patient outcomes, there's still room for improvement in predicting the exact timing of critical events.

Balanced Bias and Variance: The new model achieves a delicate balance, maintaining bias at a comparable level to the previous release while significantly enhancing the correlation performance. This balance is crucial for maintaining robust generalization across different datasets.

Conclusion

Our latest ALS model release represents a significant step forward in forecasting patient trajectories for clinical trials. By incorporating new data, features, and dependencies, and improving discriminative performance, this improved model gives researchers a more powerful tool for forecasting patient outcomes and accelerating the development of effective ALS treatments. We are excited about the potential of this model to make a meaningful impact on ALS research and improve the lives of patients. 

The specification sheet for ALS DTG 3.0 is available to download on our website and provides more information on model performance related to specific clinical trial endpoints.