Machine Learning Helps to Predict Freezing Gait in Early Parkinson’s
Changes in brain, clinic and lab data estimate risk in treatment-naive patients
Analyses done by machine learning of data collected in brain imaging, clinical exams, and laboratory tests can be used to predict the risk of freezing of gait developing among people in the early stages of Parkinson’s disease, a study reported.
The study, “Baseline cerebral structural morphology predict freezing of gait in early drug-naïve Parkinson’s disease,” was published in npj Parkinson’s Disease.
Freezing of gait, or FOG, is a common Parkinson’s motor symptom that usually develops in later disease stages. FOG is characterized by the inability to walk despite the intention to, as if one’s feet are stuck to the floor. This symptom can cause distress and inconvenience in day-to-day life, and it increases the risk of falls.
There is a lot of variability in how Parkinson’s manifests and progresses, which has historically made it difficult to predict the disease’s likely course for someone in its early stages.
Structural changes in small brain regions link strongly with freezing gait
“Recent studies have suggested that clinical assessments, laboratory tests, and brain imaging of early PD [Parkinson’s disease] patients could predict the progression” of motor problems, including FOG, the researchers wrote.
“However, these studies have only focused on a certain brain area or have a small sample” of patients with data, they added. Further, none used machine learning with that combination of data to predict FOG.
Machine learning is a form of artificial intelligence that basically uses algorithms to analyze data, learn from its analyses, and then make a prediction about something.
A team of researchers in China fed a machine learning model with clinical, laboratory, and brain imaging data from early Parkinson’s patients, as well as information about whether they developed FOG over follow-up periods. This way, it “learned” to identify patterns in these data that were associated with a future FOG risk.
“We used machine learning algorithms that are simple, easy to implement, and highly interpretable,” the researchers wrote.
The study included data for 158 adults with early Parkinson’s who were not using any disease-appropriate treatment and 73 healthy people, collected as part of the Parkinson’s Progression Markers Initiative (PPMI) study.
Over five years of follow-up, 66 of the Parkinson’s patients (49 men and 17 women) developed FOG after a median of 29 months (nearly 2.5 years). The researchers noted that men appeared more likely than women to develop this motor symptom.
Patients who ended up developing FOG had at the study’s start a poorer sense of smell, depressive symptoms, a more severe degree of the disease, and greater motor problems, like postural instability and gait difficulties, than those who did not.
Five different machine learning algorithms were tested, and results showed they could predict the risk of FOG with an accuracy of up to 78%. The researchers noted better accuracy was achieved when combining brain imaging data with clinical and laboratory findings than when using only one or the other.
Further analyses revealed that structural differences in a number of small brain regions were among the most powerful predictors of FOG in these models.
“Several disrupted brain regions … might help predict future FOG,” the researchers wrote.
These data suggest brain imaging analyses “have the potential to help predict future FOG in patients with early [Parkinson’s] at an individual level, which has higher predictive performance combined with clinical investigations,” the researchers wrote.
They noted several limitations of their study, including a relatively small number of patients and limited access to clinical data. They also noted that most participants in the PPMI database are American or European, potentially limiting a generalization of the results.