Do DTI features add value to clinical and SPECT imaging features for outcome prediction in Parkinson’s disease?

Purpose: Diffusion tensor imaging (DTI) features have shown correlation with motor outcome of Parkinson’s disease (PD) patients in previous studies. Our study sets to investigate the effect of DTI features on outcome prediction by employing hybrid machine learning (ML) systems (HMLSs). HMLSs are useful since dimensionality reduction algorithms (DRAs) linked with predictor algorithms have the potential to improve assessment of PD (including prediction of disease outcome that has the potential to significantly improve powering of clinical trials for novel disease-modifying therapies).

Methods: We selected 129 PD subjects as derived from the Parkinson’s Progressive Marker Initiative dataset (years 0 and 4) and investigated 100 features, derived from baseline (year 0). The features belonged to 3 categories: i) Non-imaging (NI) including multiple Movement Disorder Society’s Unified Parkinson's Disease Rating Scale (MDS-UPDRS) measures and a range of task/exam performances, ii) dopamine transporter (DAT) SPECT based features, including left and right putamen and caudate uptake (and their combinations), and iii) DTI based features, consisting of 3 eigenvalues as well as fractional anisotropy from left and right rostral, middle, and caudal areas of the substantia nigra. We generated 7 datasets by employing all possible combination of the 3 categories. MDS-UPDRS III in year 4 (range [3 86]) was considered as the outcome. In our study, two types of HMLSs were utilized including: i) 4 feature selection algorithms (FSAs) followed by prediction algorithms, and ii) 5 feature extraction algorithms (FEAs), both followed by prediction algorithms. The 7 prediction algorithms were regression-based and were also studied individually employing no DRAs (i.e. neither FSAs nor FEAs). The predictor algorithms underwent 5-fold cross-validation, and within the training set the algorithms was optimized via automated ML hyperparameter tuning using 15% of the training data. Predictive performance was evaluated using Mean Absolute Error (MAE).

Results: To reduce overfitting and enable improved prediction, the most significant features and attributes from the datasets were derived using FSAs and FEAs. After inputting the optimized combinations from these DRAs into multiple regressors, the smallest MAE achieved was 9.8 ± 2.0 via application of the HMLS: LASSO (least absolute shrinkage and selection operator) + MLP_BP (Multilayer Perceptron-Back Propagation) to the dataset with DAT+NI features. The smallest MAE achieved with DTI was 9.9 ± 2.1 with the DTI+DAT+NI dataset via application of HMLS: LassoGLM (least absolute shrinkage and selection operator for generalized linear models) + Linear as well as Logistic regression. The HMLSs with FEAs performed well achieving best MAE score of ~11 across all datasets. The worst performing HMLS had MAE of 39.1 ± 11.6 which consisted of no DRAs, using LOLIMOT regression algorithm, applied to the DTI+DAT+NI dataset. Overall, no enhancement of MAE was found in datasets where DTI was included vs. excluded. Conclusion: Even though DTI features were shown to be correlated with motor outcome, our study suggests that above-mentioned DTI features in year 0 do not add value to prediction of motor score (MDS-UPDRS III) in year 4, beyond usage of non-imaging and DAT SPECT based features. Furthermore, employing optimal HMLSs and multiple datasets in year 0, even using DTI, did not enable us to provide a performance compared to our previous studies resulting in a MAE ~ 4 using data from both years 0 and 1. Future work includes application of extensive radiomics features for further analysis.

• Download figure
• Open in new tab
• Download powerpoint