Routine data help AI predict Parkinson’s decline years ahead

Artificial intelligence could help identify people with Parkinson’s disease who are at increased risk of rapid cognitive or motor decline three to five years before that deterioration occurs. A study from the University of Miami found that machine learning models can distinguish between patients with different progression risks, with some of the most valuable information coming from measurements already collected during routine neurological care.

The findings could eventually support more personalised monitoring and improve the selection of participants for clinical trials. Notably, adding structural MRI measurements provided relatively little additional predictive value.

Machine learning models

Parkinson’s disease can progress very differently between individuals. Some patients remain relatively stable for years, while others experience rapid deterioration in movement, cognition or both. The researchers wanted to determine whether AI could combine different types of information to predict these trajectories. The team developed its machine learning models using data from 1,602 participants in the Parkinson’s Progression Markers Initiative. The models were subsequently tested in an independent cohort of 541 patients from the Parkinson’s Disease Biomarkers Program.

Researchers focused on two outcomes. Rapid cognitive decline was defined as a decrease of at least five points on the Montreal Cognitive Assessment, or MoCA. Rapid motor deterioration was defined as an increase of ten points on the motor component of the Movement Disorder Society Unified Parkinson Disease Rating Scale. The models combined baseline clinical information with MRI-derived measurements of changes in brain structure. Researchers also examined whether information collected during the first year after diagnosis could improve longer-term predictions.

Clinical measurements outperform MRI data

The team initially expected structural MRI measurements of brain atrophy to improve the models. Instead, routine clinical information proved considerably more valuable. The strongest clinical models achieved AUROC values above 0.80, indicating a good ability to distinguish between patients at higher and lower risk of subsequent decline. Adding structural MRI data generally did not improve performance. For some models predicting motor deterioration, it even reduced predictive accuracy.

According to senior author Ihtsham ul Haq, professor of neurology at the University of Miami Miller School of Medicine, the findings demonstrate that more data do not necessarily result in better predictions. The relevance of the information being analysed is more important than its technological sophistication. The study also highlights the potential value of longitudinal monitoring. Models became more accurate when information about how a patient’s condition changed during the first year was added to baseline measurements.

Different signals predict different outcomes

The most informative factors differed between cognitive and motor deterioration. For rapid motor decline, important predictors included results from a synuclein seed amplification assay, which detects abnormal alpha-synuclein biology associated with Parkinson’s disease. The rate at which a patient’s motor score changed during the first year after diagnosis was another major predictor.

For cognitive decline, early deterioration in cognitive performance was particularly informative. The pace of motor decline during the first year also contributed to predicting later cognitive deterioration. These findings suggest that carefully tracking relatively straightforward clinical measurements over time can reveal patterns that are difficult to identify during individual consultations. The researchers emphasise that AI is intended to support rather than replace clinical judgement.

One potential benefit is earlier identification of patients who may require closer monitoring. The researchers also point to factors such as physical activity, blood pressure, hearing loss and vision problems that may influence long-term brain health. A clearer understanding of an individual patient’s risk could help clinicians emphasise potentially relevant interventions.

Potential role in clinical trials

Predicting disease progression could also have implications for Parkinson’s research. Clinical trials investigating treatments intended to slow progression must account for substantial differences in how quickly participants naturally deteriorate.

AI models could potentially help identify people who are more likely to show measurable progression during a trial. This approach, known as trial enrichment, could make it easier to determine whether an experimental treatment is actually changing the disease trajectory rather than reflecting differences in the underlying progression rates of participants.

An important aspect of the study was its external validation. After developing the models using one large Parkinson’s dataset, the researchers tested them in an entirely separate cohort and found similar performance. This provides evidence that the identified patterns were not limited to the original study population.

Beyond structural brain imaging

The models are not presented as ready-made tools for determining an individual patient’s future. Instead, the research demonstrates how machine learning could extract prognostic information from clinical data already routinely collected in Parkinson’s care.

The researchers now plan to investigate whether similar approaches can predict rapid deterioration in other neurodegenerative conditions, including Alzheimer’s disease. They also want to determine whether measurements of brain connectivity provide additional predictive information beyond structural MRI.

For Parkinson’s care, however, one of the study’s most notable findings is relatively simple: sophisticated imaging is not necessarily required to generate useful predictions. Routine clinical assessments, particularly when repeated over time, may contain much of the information AI needs to identify patients at increased risk of rapid decline.