A statistical modeling framework based on multiclass likelihood contrast classification can accurately identify mathematical trajectory phenotypes from sparse and irregular longitudinal clinical records, according to a study evaluating Parkinson’s disease motor assessment data. In clinical research, patient measurements are frequently collected at unaligned intervals and across irregular visits, presenting significant hurdles for standard classification systems that typically depend on data interpolation or extensive feature reduction. To address these operational challenges, researchers designed an interpretable multiclass likelihood contrast framework for classifying unaligned longitudinal trajectories directly rather than compressing individual patient histories into a restricted set of summary features.
The analytical framework, developed by researchers Sangam Pangeni, Md Rejuan Haque, and Fahad Mostafa from The University of New Mexico Gallup Campus, The Ohio State University Wexner Medical Center, and Arizona State University, relies on class-specific mixed-effects models. Under this formulation, the proposed approach fits class-specific models and classifies each held-out subject according to the largest marginal likelihood. The underlying statistical specification incorporates a quadratic mean trajectory across time while allowing for subject-specific random intercepts and random slopes to capture individual baseline differences and distinct rates of trajectory change.
To evaluate how effectively the method performs under challenging data constraints, the research team conducted both numerical simulation experiments and an empirical evaluation using Parkinson’s disease motor-UPDRS data. Within the Parkinson’s disease experiment, dense longitudinal histories were initially utilized to define three mathematical trajectory phenotypes, providing an objective baseline against which the recovery of trajectory patterns could be systematically tested.
To simulate severe real-world data missingness and test trajectory recovery, the investigators deliberately masked approximately ninety percent of the longitudinal observations from the patient histories. This substantial sparsification evaluated whether the multiclass likelihood contrast algorithm could reliably recover an individual’s trajectory phenotype when provided with only sparse, unaligned records.
Model performance was assessed using subject-level cross-validation and compared directly against an extensive suite of established machine learning and statistical classifiers. Specifically, the evaluation benchmarked the proposed framework against linear mixed-likelihood classification, functional k-nearest neighbors, random forest, support vector machine, polynomial-feature classification, and grid-based principal component analysis classifiers. By testing across these diverse alternative modeling approaches, the researchers determined how effectively the likelihood contrast formulation preserved temporal structure compared to conventional feature-based methods.
Despite the substantial ninety percent sparsification of the data, the proposed model achieved an accuracy of 0.8095, a macro-F1 score of 0.8160, a Matthews correlation coefficient of 0.7107, and a macro-area under the receiver operating characteristic curve of 0.9268. Furthermore, the model exhibited consistent classification performance across different data splits, recording a validation macro-F1 score of 0.7980 alongside a held-out test macro-F1 score of 0.8160, indicating stable generalization without overfitting to specific partitions.
These findings demonstrate that multiclass likelihood contrast learning can preserve informative temporal structure in sparse and irregular longitudinal records without reducing individual trajectories to a small number of summary features. However, the authors emphasized that the Parkinson’s disease findings represent the recovery of data-derived mathematical trajectory phenotypes rather than prospective clinical prediction or diagnostic validation. The research team concluded that further evaluation in larger independent cohorts with clinically defined outcomes and multimodal longitudinal biomarkers is needed to establish genuine clinical utility.
Subject of Research: Multiclass likelihood contrast classification of sparse longitudinal clinical trajectories
Article Title: Parkinson's disease phenotype recovery from sparse longitudinal trajectories using multiclass likelihood contrast classification
Article References: Pangeni, S., Haque, M. R., & Mostafa, F. (2026). Parkinson's disease phenotype recovery from sparse longitudinal trajectories using multiclass likelihood contrast classification. Frontiers in Artificial Intelligence, 9, Article 1968393. https://doi.org/10.3389/frai.2026.1968393
Image Credits: AI Generated
DOI: 10.3389/frai.2026.1968393
Keywords: multiclass likelihood contrast, longitudinal trajectories, Parkinson's disease, motor-UPDRS, sparse longitudinal records, trajectory phenotyping, mixed-effects models, quadratic mean trajectory, random intercepts and slopes, subject-level cross-validation, linear mixed-likelihood classification, functional k-nearest neighbors, random forest, support vector machine, polynomial-feature classification, principal component analysis, Matthews correlation coefficient
News Source: Diana Fleming. (November 18, 2026). Likelihood Contrast Model Recovers Parkinson’s Trajectory Phenotypes From Sparse Data. Scienmag.



