- A machine learning study at Weill Cornell Medicine has been able to classify Parkinson’s disease into three subgroups, an advance that could help target patients with treatments specific to their disease progression.
- Analyzing data from an existing study, the researchers divided the cohort into rapid-onset, progressive-onset, and moderate-onset, an approach that recognizes the heterogeneous nature of the disease.
- Experts say the results are logical and very promising, but caution that larger populations need to be explored to create more accurate models.
On the heels of new research from Boston University showing that an artificial intelligence model could predict a person’s chances of developing Alzheimer’s disease, Weill Cornell Medicine researchers have been able to classify Parkinson’s disease into three subtypes using machine learning.
The results — which appear in
Cornell researchers analyzed data from 406 people who participated in the Parkinson’s Progression Markers Initiative (PPMI), an international observational study that “systematically collected clinical, biological, multi-omics, and brain imaging data from participants.”
They developed a deep learning model called Deep Phenotypic Progression Embedding (DPPE), which can “holistically” model “participants’ multidimensional longitudinal progression data,” as the authors explain in the study paper.
The authors also note that in recent years, Parkinson’s disease has increasingly been observed as a disease with heterogeneous symptoms and progression.
In other words, not everyone with Parkinson’s will have the same experience, and therefore treatment could be much more tailored to the needs of different patients.
The three subgroups of Parkinson’s disease identified by machine learning are based on the rate of disease progression:
- Rapid rhythm (PD-R), characterized by rapid progression of symptoms. In the observed cohort, 54 individuals (13.3%) had this subtype.
- Progressive rhythm (PD-I), which presents with mild baseline symptoms and relatively moderate progression. In the observed cohort, 145 individuals (35.7%) presented with this variety.
- Moderate-rhythm (PD-M), characterized by mild initial symptoms and moderate progression. This was the largest portion of the observed cohort, with 207 people (50.9%) living with this form of Parkinson’s disease.
The study authors note that their classifications “highlighted the need to treat (Parkinson’s disease) subtypes as unique subdisorders in clinical practice, where our rhythm subtypes could inform patient stratification and management.”
By identifying specific varieties of the disease, clinical approaches could be much more targeted and effective.
Clemens Scherzer, MD, physician-scientist and the Stephen & Denise Adams Professor of Neurology at Yale School of Medicine, who was not involved in the study, said: Today’s Medical News The study’s computational results were very interesting, but he cautioned that they were extremely preliminary and that larger populations were needed to develop and validate such classifiers.
“The goal of precision medicine is to predict the course of a patient’s disease and to intervene therapeutically upstream to prevent complications from developing. To achieve this, we need to identify the triggering factor of the disease in each patient and develop targeted therapies,” Scherzer emphasized.
“For example, we found that 10% of Parkinson’s disease patients in the United States have
Still, Dr. Daniel Truong, a neurologist and medical director of the Truong Neuroscience Institute at MemorialCare Orange Coast Medical Center in Fountain Valley, Calif., and editor-in-chief of the Journal of Clinical Parkinsonism and Related Disorders, who was also not involved in the study, said: MNT that subgroups constitute a logical and systematic approach to the treatment of Parkinson’s disease.
“For example, patients with the PD-R subtype may benefit from more aggressive treatment strategies and closer monitoring than those with the PD-I subtype, who may require less intensive management. Knowledge of a patient’s subtype can guide drug selection, including the potential repurposing of existing drugs such as metformin, which the study suggests may be particularly beneficial for the PD-R subtype.”
–Daniel Truong, MD
“This allows for predictive and preventive health care to be designed for each subtype,” Truong explained.
“Early intervention may be necessary in patients whose disease progresses rapidly. This is essential to manage symptoms before they become severe and disabling. Subtyping helps stratify patients according to their risk, which allows for more targeted and efficient clinical trials for new treatments, as well as better allocation of healthcare resources,” he added.
Steven Allder, BMedSci, BMBS, FRCP, DM, a consultant neurologist at Re:Cognition Health who was not involved in the study, agreed that identifying different subgroups in advance would allow healthcare professionals to work on specific treatment plans for each.
He listed possible treatments for each, noting:
- Progressive rhythm (PD-I): “Treatments could focus on maintaining quality of life and preventing symptom progression through lifestyle modifications, physical therapy, and possibly neuroprotective medications.”
- Moderate-Rate (PD-M): “These patients have moderate disease progression. They may benefit from a combination of pharmacological treatments to manage symptoms and slow progression, such as dopamine agonists, MAO-B inhibitors, or other disease-modifying therapies.”
- Rapid-onset (PD-R): “This subtype progresses rapidly and often involves cognitive deficits. Metformin has shown promise in improving symptoms in this group, particularly those related to cognition and falls. Early intervention with metformin and other neuroprotective agents may be crucial for the management of this subtype.”
Allder’s main concern about using machine learning technology to predict diseases like Parkinson’s was whether such a tool would be accessible to the people who need it.
“I don’t foresee any problems with the AI model, but I do foresee problems with patient access to it,” he told us.
“While AI models are powerful tools for identifying disease subtypes and predicting disease progression, there are potential challenges related to patient access. Not all patients have access to advanced diagnostic tools or treatments that come from AI research, particularly in underserved settings,” Allder said.
However, he said another problem could be “the use of large patient data for training AI models,” which “raises concerns about data privacy and security.”
“AI models need to be validated across diverse populations to ensure they are not biased toward specific cohorts,” Allder said.
Scherzer, echoing his previous statement, said the meaningful power of artificial intelligence for precise medical treatments will ultimately depend on more research and testing.
“The success of AI in predicting outcomes depends on the size and quality of the input data,” he noted. “A major gap in this field is that we need much larger, high-quality longitudinal datasets of Parkinson’s patients – data on large populations spanning the prodromal stages and the full course of the disease. These data will be essential for training and validating AI models useful for augmented medicine.”