Data-driven clinical decision support tool for diagnosing mild cognitive impairment in Parkinson's disease

Abstract

Parkinson's disease (PD) is a neurodegenerative condition that may affect both motor and cognitive function. Mild cognitive impairment (MCI) is a known risk factor for the progression to dementia in the later stages of the disease. Lengthy and time-consuming neuropsychological assessments, by trained experts, often make MCI diagnosis impractical in routine care. In this context, machine learning (ML) may offer promising support for MCI diagnosis. Thus, we analysed longitudinal data from 115 people with Parkinson's disease (PwPD) and 226 healthy control participants from the Luxembourg Parkinson's Study, combining ML with clinical data to support MCI diagnosis in PwPD. The data-driven model showed a non-inferior performance to the clinical diagnostic reference test (MDS PD-MCI Level II) and identified a subgroup of MCI individuals that was not captured by the clinical test. This finding suggests that ML models can complement clinical assessments, by facilitating the detection of MCI and complementing the diagnostic characterisation of PwPD.

Fig. 1. Diagnostic prediction strength. Clinical diagnostic reference test vs data-driven model.

Forest plot displaying Cohen’s d effect sizes and 95% confidence intervals for various features assessed by the data-driven model and the clinical diagnostic reference test. A bootstrap approach was employed to assess the statistical significance, which was determined by p-values adjusted for multiple comparisons using the Benjamini–Hochberg method p. The alpha level was set at α = 0.05. The analysed features include the MoCA total score, MDS-UPDRS 1.1 and the sum of the PDQ-39 subitems 30–33.

Fig. 2. Validation of the data-driven model for MCI subgroup identification.

The figure shows the comparison of the cognitive performance of the subgroups of participants with PD. The distribution of the global cognitive function (MoCA total score) and the physician-reported cognitive impairment status (MDS-UPDRS 1.1) among the different subgroups is displayed in Panel A and B, respectively. These variables were selected as they were not included in the model training, and are expected to differ between PD-NC and PD-MCI individuals. Statistical comparisons were performed using two-tailed Mann–Whitney U tests (due to the non-normal distribution of the features), with p-values adjusted for multiple comparisons using the Benjamini–Hochberg procedure. The alpha level was set at α = 0.05. Outliers are represented as individual diamonds, and the interquartile ranges are shown within each box.

Fig. 3. Validation of the data-driven model for MCI subgroup identification.

The current figure illustrates the risk of reaching cognitive impairment that affects activity daily of living across the different cognitive groups. This cognitive impairment was defined as either A) a MoCA total score smaller or equal to 21 or B) a moderate score in MDS-UPDRS 1. Individuals that reached the endpoint at the baseline visit were excluded from this analysis, resulting in different numbers of individuals in each analysed feature. The X-axis represents time since diagnosis (in years), and the Y-axis shows the log (hazard ratio). Each subgroup is represented by a trajectory and colour: NC (green), data-driven early-MCI group (yellow), and MCI (red). The following confounders were included in the Cox hazard analyses: age, years of education and score of the feature at baseline. Sex was not included as confounder as it violated the proportional hazards assumption.

Fig. 4. Cognitive impairment profile.

A Bar plot showing the proportion of individuals within each group (NC: green, data-driven early-MCI: yellow, MCI: red) without any impairments or with an impairment by cognitive domain (attention, executive, memory, visuospatial, and language). B Bar plot showing the proportion of individuals without impairment (green), single domain impairment (yellow), and multiple domain impairment (red) within the NC, data-driven early-MCI, and MCI groups. Percentages, in both graphs, were calculated relative to the total number of participants in each group (NC: n = 49; data-driven early-MCI: n = 26; MCI: n = 39). C Box plot showing the differences in depression and subjective cognitive decline across the identified groups. Each subgroup is represented by a colour: NC (green), data-driven early-MCI group (yellow), and MCI (red). Statistical comparisons were performed using two-tailed Mann–Whitney U tests (due to the non-normal distribution of the features), with p-values adjusted for multiple comparisons using the Benjamini–Hochberg procedure. The alpha level was set at α = 0.05. Outliers are represented as individual diamonds, and the interquartile ranges are shown within each box.