IBM and MIT Researchers Introduce a New Machine Learning Approach that Finds Nonlinear Patterns of Neurodegenerative Disease Progression

Amyotrophic lateral sclerosis (ALS), Alzheimer’s disease, and Parkinson’s disease are all examples of neurodegenerative disorders, which are complex and chronic illnesses. These disorders have a wide range of symptoms, varying progression rates, and various genetic and environmental causes, some of which are still unknown. Knowing how a disease like ALS develops is essential for enrolling patients in clinical trials, evaluating treatment options, and identifying causes.

However, disease evolution assessment is not that simple. To determine if a medicine is effective in delaying the progression of an illness, linear models are commonly used in current clinical research, which assumes that health diminishes linearly on a symptom rating scale. Data, however, suggest that ALS progression is frequently nonlinear, with periods of stable symptoms interspersed with periods of fast change. In addition, comparisons across patient populations can be problematic due to data scarcity and the fact that many health assessments rely on subjective rating metrics obtained at varying periods. Because of this variation, it is more difficult to assess the impact of inventions and may even obscure the actual cause of diseases.

Researchers from MIT and IBM Research have created a new machine-learning approach to define ALS disease progression patterns for use in designing treatment trials. Their method successfully revealed distinct and robust clinical patterns in the development of ALS, many of which defy linear explanation. In addition, these subgroups of disease progression were similar across patient demographics and illness measures. 

Gaussian process regression and Dirichlet process clustering made up the basis of the unsupervised machine learning model they developed. These automatically grouped health trajectories with similar trajectories without mandating the number of clusters or the shape of the curves, generating ALS progression “subtypes” estimated from patient data. Their method did not presume linearity but did incorporate existing clinical knowledge in the form of a bias toward negative trajectories, which is in line with expectations for the course of neurodegenerative diseases. The used approaches and models were more adaptable because “they capture what was seen in the data,” avoiding the requirement for costly labeled data and prescription of parameters.

Five longitudinal datasets from ALS clinical trials and observational research were used as the primary data sources for the model. They used  ALS functional rating scale revised (ALSFRS-R) and also integrated performance on survivorship probability, forced vital capacity (a measure of respiratory function), and subscores of the ALS Functional Rating Scale-Revised (ALSFRS-R), which evaluates specific body functions.

The model was trained and tested on these metrics for the entire population, and four distinct patterns of illness progression emerged: sigmoidal fast progression, stable slow progression, unstable slow progression, and unstable moderate progression. Notably, it detected trajectories in which patients suffered an abrupt loss of ability, known as a functional cliff, which would have a major effect on treatments, clinical trial enrolment, and quality of life.

To disentangle the effects of clustering and linearity on model accuracy, the researchers compared their methodology to well-used linear and nonlinear methods in the field. The new approach outperformed the others, including patient-specific models, and discovered similar patterns of subtypes across several evaluation metrics. Surprisingly, the model could interpolate missing values and significantly predict future health measurements, even when data were withheld. The model was resilient, generalizable, and accurate even with limited data because it could be trained on a single ALSFRS-R dataset and then used to predict cluster membership in others. 

This study presents substantial progress toward deciphering the time series of complicated neurodegenerative illnesses. The researchers claim that the patterns they’ve seen are consistent across investigations, which could have practical ramifications for their classification of disease [ALS]. This method also sheds light on Alzheimer’s and Parkinson’s disorders, whose symptoms and course can vary widely.

This Article is written as a research summary article by Marktechpost Staff based on the research paper 'Machine learning approach finds nonlinear patterns of neurodegenerative disease progression'. All Credit For This Research Goes To Researchers on This Project. Check out the paper and reference article.

Please Don't Forget To Join Our ML Subreddit

Tanushree Shenwai is a consulting intern at MarktechPost. She is currently pursuing her B.Tech from the Indian Institute of Technology(IIT), Bhubaneswar. She is a Data Science enthusiast and has a keen interest in the scope of application of artificial intelligence in various fields. She is passionate about exploring the new advancements in technologies and their real-life application.