Detection of mild cognitive impairment in Parkinson’s disease using gradient boosting decision tree models based on multilevel DTI indices

Abstract

Background Cognitive dysfunction is the most common non-motor symptom in Parkinson’s disease (PD), and timely detection of a slight cognitive decline is crucial for early treatment and prevention of dementia. This study aimed to build a machine learning model based on intra- and/or intervoxel metrics extracted from diffusion tensor imaging (DTI) to automatically classify PD patients without dementia into mild cognitive impairment (PD-MCI) and normal cognition (PD-NC) groups. Methods We enrolled PD patients without dementia (52 PD-NC and 68 PD-MCI subtypes) who were assigned to the training and test datasets in an 8:2 ratio. Four intravoxel metrics, including fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD), and two novel intervoxel metrics, local diffusion homogeneity (LDH) using Spearman’s rank correlation coefficient (LDHs) and Kendall’s coefficient concordance (LDHk), were extracted from the DTI data. Decision tree, random forest, and eXtreme gradient boosting (XGBoost) models based on individual and combined indices were built for classification, and model performance was assessed and compared via the area under the receiver operating characteristic curve (AUC). Finally, feature importance was evaluated using SHapley Additive exPlanation (SHAP) values. Results The XGBoost model based on a combination of the intra- and intervoxel indices achieved the best classification performance, with an accuracy of 91.67%, sensitivity of 92.86%, and AUC of 0.94 in the test dataset. SHAP analysis showed that the LDH of the brainstem and MD of the right cingulum (hippocampus) were important features. Conclusions More comprehensive information on white matter changes can be obtained by combining intra- and intervoxel DTI indices, improving classification accuracy. Furthermore, machine learning methods based on DTI indices can be used as alternatives for the automatic identification of PD-MCI at the individual level.

Machine learning for the detection and diagnosis of cognitive impairment in Parkinson’s Disease: A systematic review

“…A number of study activities were utilised including diagnosis, (PD vs Healthy Controls (HC), PDD vs HC, PD-NC vs PD-MCI, PD-MCI vs PDD), differential diagnosis (PD-CI/PD-MCI/PDD vs AD vs DLB), identification of biomarkers for PD detection, and the prediction of future CI states. Most studies focused on diagnostic activities (n = 48) [ 21 – 26 , 28 , 30 , 31 , 84 , 113 , 114 , 116 – 119 , 121 , 122 , 124 – 126 , 128 , 130 , 133 – 136 , 138 , 142 , 144 – 146 , 148 – 151 , 153 – 155 , 157 – 161 , 164 , 166 , 171 , 172 ], followed by prediction (n = 12) [ 29 , 129 , 131 , 132 , 137 , 140 , 141 , 152 , 156 , 162 , 163 , 170 ], biomarker identification (n = 6) [ 27 , 120 , 123 , 143 , 147 , 169 ], and differential diagnosis (n = 4) [...…”

“…The most commonly used data modalities were imaging (n = 33) [ 24 , 27 – 31 , 113 , 114 , 117 , 119 , 122 , 129 , 130 , 133 , 137 , 139 – 143 , 147 – 152 , 155 , 156 , 159 – 161 , 163 , 164 ], clinical characteristics (n = 17) [ 84 , 118 , 131 , 132 , 137 , 140 , 152 , 153 , 155 , 156 , 159 , 160 , 162 – 164 , 169 , 170 ], EEG (n = 11) [ 25 , 26 , 120 , 125 – 128 , 135 , 136 , 146 , 151 ], and neuropsychological profile (n = 10) [ 115 , 118 , 132 , 134 , 142 , 157 , 158 …”

“…ML techniques used across all reviewed studies were categorised into 12 categories, some of which overlap: (1) tree based methods (n = 32) [ 22 , 23 , 26 , 27 , 31 , 114 , 116 , 119 , 122 , 125 , 126 , 128 – 130 , 132 , 134 , 137 , 139 , 141 , 145 , 153 , 155 , 157 – 159 , 162 , 164 , 166 , 167 , 170 – 172 ], (2) Support Vector Machines (n = 30) [ 21 , 23 – 25 , 27 , 28 , 30 , 113 , 115 , 117 , 118 , 122 – 124 , 133 , 134 , 138 , 139 , 141 , 142 , 148 – 151 , 153 , 156 – 159 , 161 , …”

“…The most commonly used metric was accuracy (n = 46), used both as a sole performance metric (n = 6) [ 25 , 121 , 123 , 154 , 155 , 166 ] and as part of a combination with other metrics (n = 40). In studies using a combination of metrics, the most common combination was accuracy, sensitivity, and specificity (n = 17) [ 23 , 113 , 114 , 116 , 117 , 124 – 126 , 130 , 133 , 142 , 149 , 150 , 158 , 164 , 167 , 171 ], alongside accuracy, sensitivity, specificity, and Area under the ROC Curve (AUC) (n = 13) [ 22 , 27 , 31 , 115 , 119 , 127 , 135 – 137 , 148 , 151 , 153 , 161 ].…”

“…44 of the 58 supervised learning studies focused on classification activities, 40 of which focused on classification vs NC/HC [ 22 – 26 , 28 , 30 , 31 , 84 , 113 , 114 , 116 , 117 , 119 , 122 , 124 – 126 , 128 , 130 , 133 , 134 , 138 , 142 , 145 , 146 , 148 – 151 , 153 – 155 , 158 , 159 , 161 , 164 , 166 , 171 , 172 ], and four classifying against memory disorders [ 115 , 127 , 139 , 167 ]. 11 studies focused on prediction [ 129 , 131 , 132 , 137 , 140 , 141 , 152 , 156 , 162 , 163 , 170 ], and the remaining three on biomarker identification [ 27 , 123 , 169 ].…”