HCP Live
Contagion LiveCGT LiveNeurology LiveHCP LiveOncology LiveContemporary PediatricsContemporary OBGYNEndocrinology NetworkPractical CardiologyRheumatology Netowrk

Machine Learning Helps Identify Parkinson Disease States, Disability Progression

The model discovered nonsequential, overlapping disease progression trajectories, supporting the use of nondeterministic disease progression models.

Researchers from the Michael J. Fox Foundation and IBM recently published data on a newly developed statistical progression model that identifies 8 Parkinson disease (PD) states and categorizes patients based on differences in motor and nonmotor symptoms and disease progression.

Lead author Kristen A. Severson, PhD, Center for Computational Health, IBM, and colleagues collected data on 423 patients with early PD and 196 health controls (HCs) for up to 7 years from the Parkinson’s Progression Markers Initiative (PPMI). A contrastive latent variable model was applied followed by a novel personalized input-output hidden Markov model to define disease states.

"With validation of the model in real-world settings with diverse cohorts capturing a range of clinically relevant short-term and long-term outcomes, and further refinement to leverage a minimal set of emerging clinical or biomarker assessments that best identify granular states of Parkinson's disease progression, our model has the potential to translate to a prognostication tool in the clinic, and a predictive tool that might be used for clinical trial sample enrichment,” Severson et al wrote.

Additionally, Severson and colleagues noted that the model accounts for the effects of Parkinson's disease medications—something prior models have considered in a limited fashion or not at all. "We modelled medication effects as deviations from disease states, which vary across states and participants. An ideal model might allow these factors to vary together, with individualized response varying over time. However, most studies do not have enough data to accurately estimate time-varying, individualized response. Therefore, we compromise by allowing each patient to have a non-time-varying personalized response to medication and a time-varying response which is a function of disease state," they wrote.

READ MORE: Novel PET Imaging Tracer Shows Promise for Parkinson Disease Diagnosis

Seven clinically relevant key outcomes that did not contribute to the discovery model were defined to test the clinical relevance of the identified states: mild cognitive impairment (MCI), dementia, dyskinesia, motor fluctuation, functional impairment from motor fluctuations, Hoehn and Yahr score greater than 3, and death. The investigators validated their results using an independent sample of 610 patients with PD from the National Institute of Neurological Disorders and Stroke Parkinson’s Disease Biomarker Program (PDBP).

States were indexed starting from state 1, which was mildest, to state 8, which was considered terminal. There were multiple agreements between the model-designed state and the clinical observations for individuals assigned to that state. For instance, states 5 and 8 were associated with greater postural instability or gait difficulty, and most pairwise comparisons of states in these measures were assessed by Movement Disorders Society–Unified Parkinson Disease Rating Scale (MDS-UPDRS) P values less than .01.

In the PPMI cohort, state 8 accounted for 94 (56%) of 169 MCI cases, 18 (95%) of 19 dementia cases, 5 (71%) of 7 dyskinesia cases, 22 (55%) of 40 motor fluctuation cases, 58 (68%) of 85 functionally impactful motor fluctuation cases, 28 (90%) of 31 cases with Hoehn and Yahr scores greater than 3, and 10 (56%) of 18 instances of death. In the PDBP cohort, state 8 accounted for 9 (82%) of 11 dyskinesia cases, 13 (54%) of 24 motor fluctuation cases, 162 (74%) of 219 functionally impactful motor fluctuation cases, and 65 (97%) of 67 cases with Hoehn and Yahr scores greater than 3.

At the outset of the PPMI cohort, state 1 was the most observed state, accounting for 96 of 233 participants (29%) in the PPMI-training set and 21 of 83 participants (25%) in the PPMI-testing set. In total, 4 of 333 (1%) PPMI-training participants and 3 of 83 (4%) PPMI-testing participants were in state 8 at the study onset. By year 5, 41% and 58% of patients in each group, respectively, reached state 8. Although, Kaplan-Meier analysis showed that the ranking of the starting state did not match the ranking of reaching state 8 within 5 years.

Throughout the entire study period, the most observed state transitions were from 7 to 8, 6 to 7, and 4 to 6. Notably, no state transitions were observed from 1 to 8, 2 to 3, or 4 to 7. Overall, participants in state 5 had the shortest time to state 8 (PPMI-training set: median time, 2.75 years [95% CI, 1.75-4.25]; PPMI-testing set: median time, 2.25 [95% CI, 0.5-3.25]), and were significantly more likely to reach state 8 with participants starting in states 1, 3, and 4.

Severson KA, Chahine LM, Smolensky LA, et al. Discovery of Parkinson’s disease states and disease progression modelling: a longitudinal data study using machine learning. Lancet Digital Health. doi: 10.1016/S2589-7500(21)00101-1