Out of 1,602 patients and then 541 in external validation, fed models of simple clinical data exceed 0.80 AUROC to distinguish which will decline rapidly within three to five years.
A team from the University of Miami that brings together neurologists, radiologists and AI specialists — including Ihtsham ul Haq, Yelena Yesha, Yusen Wu and Tatjana Rundek — publishes machine learning models in npj Parkinson's Disease designed to identify those who are at risk of rapid cognitive or motor deterioration in Parkinson's disease within three to five years.
The training is based on 1,602 participants from the Parkinson's Progression Markers Initiative, with external validation on 541 patients from the Parkinson's Disease Biomarkers Program. The models combine basic clinical data and MRI-derived measures of brain atrophy. The purely clinical versions achieve AUROCs greater than 0.80, a robust ability to separate high-risk profiles from low-risk profiles.
The most useful result in practice concerns the hierarchy of variables. For motor decline, it is the synuclein test and the rate of degradation of the MDS-UPDRS motor score during the first year that weigh the most; for cognitive decline, the rate of early cognitive deterioration and that of motor decline. In other words, most of the signal comes from measurements that neurologists already collect in consultations. As Ihtsham ul Haq summarizes, the most useful information was not necessarily the most technologically sophisticated.
Read the full article on the original source
Go to the source →