Optimal Feature Selection and Machine Learning for Prediction of Outcome in Parkinson’s Disease

Objectives: Features selection is very important to improve machine learning (ML) based assessment of Parkinson’s diseases (PD) including prediction of disease subtypes, which can be sensitive to the choice of features used in the models. We set to investigate optimal feature selection and ML for prediction of PD outcome (disease subtype in year 4).

Methods: We selected 885 PD subjects as derived from longitudinal datasets (years 0, 1, 2 & 4; Parkinson’s Progressive Marker Initiative), and investigated 982 features including multiple Movement Disorder Society’s Unified Parkinson's Disease Rating Scale (MDS-UPDRS) measures, a range of task/exam performances, socioeconomic/family histories and SPECT image features. Segmentation of regions-of-interest (ROIs; caudate and putamen) on DaT SPECT images were performed via MRI images. Radiomic features (RFs) were extracted for each ROI using our standardized SERA software. Hybrid ML systems (HMLS) were constructed invoking: 9 feature selection algorithms (FSAs) including ILFS (Infinite Latent Feature Selection) and LASSO (least absolute shrinkage and selection operator), and 11 classifiers including KNN (K Nearest Neighborhood), Random Forest algorithm (RFA), Ensemble Learner Algorithm (ELA) and RNN (Recurrent Neural Network). We first performed elaborate unsupervised clustering to identify disease subtypes (3 clusters robustly identified in another work of ours). We then applied FSAs to the labeled cluster data, ranking features based on importance for each FSA. We then implemented clustering using sequential arrangements including combinations of the ranked features (i.e. {1}, {1,2}, {1,2,3}, {1,2,3,4}, ⋯, {1,⋯,N}). We then set a minimum threshold of 80% correlation (i.e. looking for minimum number of features reaching similar performance as using all features). To specify optimal combinations, we first split data based on years 0 and 4, and then applied those to multiple classifiers optimized via automated ML hyperparameter tuning and 5-fold cross-validation. Subtypes in year 4 were considered as outcome. The average accuracy of folds for each method were reported and compared. For specifying important features, a voting procedure was implemented via scoring features (in a range [0-1]) based on their ranks resulting in each FSA. Finally, we applied combination with high score features to multiple classifiers.

Results: The most important features (out of 982) were selected for each FSA to enable high correlations with using all features (arriving at ~300-400 features). After applying those combinations to multiple classifiers, the best accuracy of 94% was obtained via HMLS: ILFS + KNN; some HMLSs such as ILSF + RFA or ELA and LASSO + ELA also resulted in accuracy of 90%. Similar performances were achieved when using all features. Most importantly identified features were SPECT based RFs, although three clinical features including MDS-UPDRS III (motor symptom), Semantic Fluency score (non-motor symptom) and Symbol Digit Modalities Text score (non-motor symptom) were also high performers.

Conclusions: We demonstrated that an appropriate HMLS framework enabled robust prediction of disease subtypes in PD subjects. The KNN classifier was able to work with large input features directly, resulting in a prediction accuracy of 94%. In another approach, where FSAs were first utilized, pre-selecting optimal features prior to application of classifiers, a similar performance was obtained through coupling ILFS with KNN. We showed that SPECT based radiomic features are important in clustering and classifying PD subjects. Overall, we conclude that combining medical information with SPECT-based radiomic features, and optimal utilization of HMLS, can produce very good prediction of disease outcome in PD patients.

• Download figure
• Open in new tab
• Download powerpoint