AI model helps predict ALS mortality using clinical metrics
Study: It may aid clinicians, families in making treatment, end-of-life decisions
Using seven clinical measures that are fairly easy to assess, researchers have developed a machine learning model to help predict mortality in people with amyotrophic lateral sclerosis (ALS), according to a study.
The model, which was trained on data from more than 1,900 patients, may aid clinicians and families in making decisions about treatment and end-of-life care. It performed about as well as existing tools, but offers the advantage of relying on simple, widely available data from a single clinical visit that can be updated over time.
“Our approach signifies a step [toward] harnessing the power of predictive modeling to guide healthcare providers, their patients, and caregivers as they consider goals of care, advance directives, and end-of-life decisions,” researchers wrote.
The model was described in the study, “Predicting Amyotrophic Lateral Sclerosis Mortality With Machine Learning in Diverse Patient Databases,” which was published in Muscle & Nerve.
Existing survival models have limitations
ALS is a progressive disease marked by the loss of motor neurons, the nerve cells that control voluntary movement. As these cells die, muscles weaken, affecting movement, speech, swallowing, and breathing. Treatments can slow progression and extend survival in some cases, but there is no cure.
Knowing a person’s likely survival can help clinicians, patients, and families plan ahead and inform decisions about palliative care, clinical trial participation, or assistive devices such as ventilators and feeding tubes.
Several survival models exist, but these tend to rely only on information from one clinical visit, often around the time of diagnosis, which limits their ability to capture the variability of disease progression.
“Patients often make multiple clinical visits, but single-visit prediction models do not utilize this additional information and account for ALS progression variability, potentially affecting prediction accuracy,” the researchers noted.
Some of the higher-performing prognostic models also require data from MRI scans and other expensive tools, which aren’t always available.
AL model works well with only 7 variables
To address these limitations, a team of researchers in Singapore built a new model using basic clinical and laboratory data that could estimate the probability of surviving six or 12 months after a given clinic visit.
The model was trained on information from 24,020 visits by 1,941 participants included in the PRO-ACT database, which features data from hundreds of patients who took part in ALS clinical trials. It was also designed to update predictions as new patient data become available.
Results showed that the model performed slightly better than an existing model at predicting mortality after six and 12 months, either when only five features from an older model were included or when all 48 clinical, laboratory, and demographic features from PRO-ACT were examined.
When data from multiple visits over 2 to 3 months were pooled, predictions for one-year mortality became even more reliable.
The researchers also examined which measures most strongly influenced predictions. The top 10 features included laboratory markers of inflammation, nutritional status, respiratory function, changes in ALS Functional Rating Scale-Revised scores, and body mass index, a measure of body fat.
Even small improvements may be clinically meaningful in high-stakes settings like end-of-life prognostication.
When testing if any variables could be removed from the model, the team found that the model could work well with only seven variables. Individually, none of these factors could predict survival with confidence, but together, they allowed the model to work effectively. Limiting the number of variables like this could simplify its use in clinical practice.
The model was compared with the established ENCALS survival prediction algorithm. While ENCALS was easier to use, the new model showed slightly better accuracy in some analyses. These differences were subtle, but “even small improvements may be clinically meaningful in high-stakes settings like end-of-life prognostication,” the researchers wrote.
When tested on additional datasets, including data from a U.S. clinical trial and a Singapore-based observational study, both models performed similarly, supporting their robustness across different populations.
Overall, the findings suggest that machine learning — an arm of artificial intelligence that uses algorithms to enable computers to learn from data, identify patterns, and make decisions or predictions — could be applied to simple clinical measures to support personalized and timely discussions in ALS care.
About the Author
