AI Learns to Predict the Risk of Death from Parkinson's Disease

Scientists from Yonsei University (South Korea) conducted a study on predicting the mortality of patients with Parkinson's disease using explainable artificial intelligence (XAI), which is able to describe its actions and decisions. For the work, the specialists used data from 36,480 people. For individual forecasting, 165 parameters were derived from the data, including age, gender, income level, area of ​​residence, and type of insurance. AI, using seven machine learning algorithms, identified the most influential characteristics for predicting mortality. The results of the study were published in the journal Nature.

Of the algorithms, XGBoost showed the best performance. The model identified age, gender (male), pneumonia, Alzheimer's disease, and high body mass index as the most significant characteristics for predicting mortality. The markers also included ischemic infarction, unspecified dementia, and traumatic brain injury.

The table compared data from patients who died within 10 years of diagnosis with those living with the disease for the same period. In addition to medical records from the National Health Insurance Service of South Korea (NHIS), the researchers used information from personal questionnaires of patients, which helped to detail indicators of people's quality of life. The model showed that patients with low physical activity died more often (13,484 people out of 36,480 over 10 years). At the same time, only 809 people who led an active lifestyle died during this period.

Another parameter studied, fasting glucose (mg/dl), averaged 107.79 ± 35.47 in deceased patients and 103.43 ± 27.84 in survivors. Deceased patients had lower cholesterol levels (188.19 ± 45.26 versus 193.82 ± 44.59 in survivors) and body mass index (23.21 ± 3.29 versus 24.21 ± 3.10).

The scientists point out that by using XAI they were able to demonstrate the possibility of individually predicting the mortality risk in patients with Parkinson's disease at the time of diagnosis. Accordingly, the controlled parameters (BMI, laboratory parameters) can be adjusted to prolong life. The practical application of these data allows for personalized patient management: for example, elderly men with Parkinson's disease and concomitant Alzheimer's disease are recommended to have special weight and laboratory parameters monitoring. However, the researchers emphasize that in order to increase the accuracy of the ten-year forecast in future studies, it is advisable to integrate dynamic clinical parameters. Validation of the model on other populations is also necessary for scaling the model.

In September 2023, one of the winners of the annual Breakthrough Prize, informally called the “Oscars of Science,” was a research team from Germany that discovered the most common genetic causes of Parkinson’s disease. Thomas Gasser (Hertie Institute for Clinical Brain Research, University of Tübingen and German Center for Neurodegenerative Diseases), Ellen Sidransky (National Human Genome Research Institute, National Institutes of Health), and Andrew Singleton (National Institute on Aging, National Institutes of Health) discovered the most common genetic causes of Parkinson’s disease.

Sidransky identified mutations in the GBA1 gene, which codes for an enzyme that breaks down fatty substances in cells, which is a genetic risk factor for Parkinson’s disease. Meanwhile, Gasser and Singleton independently showed that mutations in the LRRK2 gene lead to increased activity of a protein thought to contribute to the neuronal damage that occurs in the disease. These findings offer clues to the mechanisms that cause the disease, pointing to the role of the lysosome, a cellular organelle that breaks down and recycles cellular components.