A statistical method recovered predefined Parkinson’s movement-score trajectories with about 81% accuracy despite extensive missing observations. This was a methodological experiment, not a diagnostic trial.
What researchers studied
The researchers developed a multiclass likelihood approach for irregular longitudinal data, where people have measurements at different intervals.
Their method fits class-specific mixed-effects trajectories and allows variation between individuals.
They evaluated it using simulated datasets and Parkinson’s disease motor-UPDRS records.
Dense measurement histories were first used to define three mathematical trajectory groups.
Approximately 90% of the observations were then masked to test whether the model could recover those groups from sparse records.
Subject-level validation was used to compare the approach with other statistical and machine-learning classifiers.
This evaluation focuses on recovery of constructed trajectory labels rather than independently diagnosed clinical subtypes.
What they found
8160 macro-F1.
9268.
8160.
These metrics suggest that useful information about temporal patterns survived substantial masking in the study dataset.
They do not indicate an 81% chance of correctly diagnosing a new patient.
What it means
Patient monitoring often produces incomplete data because appointments are missed or scheduled irregularly.
Methods that preserve the shape of measurements over time may help researchers explore progression patterns without reducing records to simple averages.
The study provides a possible tool for analysing sparse clinical datasets and comparing mathematical trajectory groups.
Further research could test whether such groupings correspond to clinically meaningful outcomes or treatment decisions.
For now, the contribution is primarily statistical rather than a validated bedside application.
Important limitation
The groups were defined mathematically from dense histories, and missingness was introduced artificially.
Real-world missing data may be systematically related to illness severity or access to care.
External clinical validation would be needed before applying the method to decisions about patients.
Source Information
Study: Parkinson’s disease phenotype recovery from sparse longitudinal trajectories using multiclass likelihood contrast classification
Authors: Sangam Pangeni, Md Rejuan Haque and Fahad Mostafa
Journal / Source: Frontiers in Artificial Intelligence
Published: 18 November 2026







