Previous studies have shown a correlation between diffusion tensor imaging (DTI) features and motor outcomes in Parkinson’s disease (PD) patients. Our study investigates the effect of DTI features on outcome prediction using hybrid machine learning systems (HMLSs).

Methods

From the Parkinson's Progressive Marker Initiative dataset (years 0 and 4), we selected 129 PD subjects and investigated 100 features derived from baseline (year 0). Three categories of features were identified: 

1) non-imaging (NI), including Movement Disorder Society's Unified Parkinson's Disease Rating Scale (MDS-UPDRS) measures and task/exam performances.

 2) SPECT-based features assessing putamen and caudate concentrations (and their combinations).

3) DTI-based features, consisting of 3 eigenvalues and fractional anisotropy from left and right rostral, middle, and caudal areas of the substantia nigra.

Using all possible combinations of the three categories, we created seven datasets. Our study used two types of HMLSs: 1) four feature selection algorithms (FSAs), followed by prediction algorithms. 2) five feature extraction algorithms (FEAs), both followed by prediction algorithms. Seven regression-based prediction algorithms were studied individually without DRAs (i.e., no FSAs or FEAs). In the training set, 15% of the training data was used to tune the ML hyperparameters of the predictor algorithms, which underwent fivefold cross-validation. Evaluation of predictive performance was done using Mean Absolute Error (MAE).

Results

Using FSAs and FEAs, the most significant features and attributes were derived from the datasets to reduce overfitting and improve prediction. The smallest MAE was obtained using the HMLS: LASSO (least absolute shrinkage and selection operator) + MLP_BP (Multilayer Perceptron-Back Propagation) method with DAT+NI features after putting the optimized combinations from these DRAs into multiple regressors. Based on the DTI+DAT+NI dataset, the smallest MAE was achieved using HMLS: LassoGLM (least absolute shrinkage and selection operator for generalized linear models) + linear as well as logistic regression. All HMLSs with FEAs achieved the best MAE score of ~11 across all datasets. HMLS with the lowest MAE received a DTI+DAT+NI dataset with no DRAs and took the LOLIMOT regression algorithm to calculate MAE. A comparison of datasets where DTI was included and excluded found no enhancement in MAE.

Conclusion

Despite the statistical evidence that those DTI features are correlated with motor outcomes in year 0, our study suggests that they do not add any value to motor score prediction (MDS-UPDRS III) in year 4, aside from non-imaging and DAT SPECT features. Further, even using a DTI, even when employing optimal HMLSs and multiple datasets in year 0, we could not achieve a performance comparable to our previous research resulting in an MAE of -4. In the future, extensive radionics features will be applied for further analysis.

Paper source: https://jnm.snmjournals.org/content/62/supplement_1/1414