Introduction
Methods
Patients
Training—validation data set | Test data set | |
---|---|---|
N | 137 | 102 |
Age | 75 ± 8 | 67 ± 15 |
Sex (male) | 64 (47%) | 43 (42%) |
PS/PD/DLB | 105 (77%) | 65 (64%) |
PD | 35 (26%) | 20 (20%) |
PS | 55 (40%) | 34 (33%) |
DLB | 6 (4%) | 11 (11%) |
PD or PS | 8 (6%) | 0 (0%) |
PD or DLB | 1 (1%) | 0 (0%) |
Non-PS/PD/DLB | 32 (23%) | 37 (36%) |
Alzheimer disease | 1 (1%) | 2 (2%) |
Essential tremor | 2 (2%) | 4 (4%) |
Other diseases and/or unknown etiology | 29 (21%) | 31 (30%) |
Image features | ||
Low | 106 (77%) | 47 (46%) |
Asymmetric | 82 (60%) | 38 (37%) |
Dot-like | 88 (64%) | 48 (47%) |
Abnormal | 108 (79%) | 42 (41%) |
123 I-ioflupane SPECT
Data processing
Image interpretation
Machine learning
ROI-based parameters
Multivariable models for a diagnosis of PS/PD/DLB
Ethics approval
Statistical analysis
Results
Fourfold cross-validation in the training data set
Features | AUC | Recall | Precision | F1 score | Accuracy | p |
---|---|---|---|---|---|---|
High or low | ||||||
LR | 0.96 | 0.94 | 0.71 | 0.81 | 0.90 | < 0.0001 |
kNN | 0.96 | 0.84 | 0.79 | 0.81 | 0.91 | < 0.0001 |
GBT | 0.95 | 0.97 | 0.61 | 0.75 | 0.85 | < 0.0001 |
SBR | 0.92 | 0.97 | 0.64 | 0.77 | 0.87 | < 0.0001 |
Symmetry or asymmetry | ||||||
LR | 0.57 | 0.49 | 0.70 | 0.81 | 0.57 | 0.577 |
kNN | 0.75 | 0.79 | 0.81 | 0.81 | 0.77 | < 0.0001 |
GBT | 0.54 | 0.18 | 0.79 | 0.81 | 0.48 | 0.059 |
Asymmetry index | 0.68 | 0.83 | 0.75 | 0.81 | 0.72 | 0.017 |
Comma or dot | ||||||
LR | 0.91 | 0.86 | 0.95 | 0.91 | 0.88 | < 0.0001 |
kNN | 0.91 | 0.88 | 0.90 | 0.89 | 0.85 | < 0.0001 |
GBT | 0.94 | 0.82 | 0.97 | 0.89 | 0.87 | < 0.0001 |
PC ratio | 0.85 | 0.74 | 0.92 | 0.82 | 0.79 | < 0.0001 |
Normal or abnormal | ||||||
LR | 0.91 | 0.86 | 0.95 | 0.91 | 0.88 | < 0.0001 |
kNN | 0.91 | 0.88 | 0.90 | 0.89 | 0.85 | < 0.0001 |
GBT | 0.94 | 0.82 | 0.97 | 0.89 | 0.87 | < 0.0001 |
Identification of features in the test data set
AUC | Recall | Precision | F1 score | Accuracy | p | |
---|---|---|---|---|---|---|
High or low | ||||||
LR | 0.95 | 1.00 | 0.67 | 0.80 | 0.85 | < 0.0001 |
kNN | 0.95 | 1.00 | 0.67 | 0.80 | 0.85 | < 0.0001 |
GBT | 0.94 | 0.90 | 0.77 | 0.83 | 0.89 | < 0.0001 |
SBR | 0.96 | 1.00 | 0.67 | 0.80 | 0.85 | < 0.0001 |
Symmetry or asymmetry | ||||||
LR | 0.58 | 0.53 | 0.76 | 0.63 | 0.63 | 0.202 |
kNN | 0.68 | 0.62 | 0.77 | 0.69 | 0.67 | 0.003 |
GBT | 0.67 | 0.57 | 0.79 | 0.66 | 0.66 | 0.001 |
Asymmetry index | 0.64 | 0.95 | 0.71 | 0.81 | 0.75 | < 0.0001 |
Comma or dot | ||||||
LR | 0.92 | 0.73 | 0.98 | 0.84 | 0.81 | < 0.0001 |
kNN | 0.79 | 0.71 | 0.89 | 0.79 | 0.76 | < 0.0001 |
GBT | 0.88 | 0.83 | 0.93 | 0.88 | 0.84 | < 0.0001 |
PC ratio | 0.81 | 0.65 | 0.88 | 0.75 | 0.72 | < 0.0001 |
Normal or abnormal | ||||||
LR | 0.91 | 0.96 | 0.58 | 0.73 | 0.81 | < 0.0001 |
kNN | 0.89 | 1.00 | 0.57 | 0.72 | 0.80 | < 0.0001 |
GBT | 0.91 | 0.89 | 0.68 | 0.77 | 0.86 | < 0.0001 |
Relationships between features and variables
Analysis of combinations of variables to identify PS/PD/DLB using AUC
Method | AUC | Sensitivity | Specificity | p | |
---|---|---|---|---|---|
ROI-based methods | |||||
Specific binding ratio (SBR) | 0.85 | 0.94 | 0.65 | < 0.0001 | |
Asymmetry index | 0.77 | 0.54 | 0.95 | < 0.0001 | |
PC ratio | 0.71 | 0.69 | 0.70 | < 0.0001 | |
SBR + asymmetry index (Model 1) | 0.86 | 0.91 | 0.73 | < 0.0001 | |
SBR + asymmetry index + PC ratio | 0.86 | 0.89 | 0.73 | < 0.0001 | |
Machine learning-based feature | |||||
Normal or abnormal (Model 2) | LR | 0.82 | 0.91 | 0.70 | < 0.0001 |
GBT | 0.88 | 0.83 | 0.87 | < 0.0001 | |
Combined model with ML features | |||||
Age + HiLo (LR) + SymAsym (kNN) + CommaDot (LR) | 0.90 | 0.79 | 0.95 | < 0.0001 | |
Best forward-stepwise model: age + HiLo(GBT) + SymAsym (kNN) + CommaDot (LR) (Model 3) | 0.93 | 0.86 | 0.92 | < 0.0001 |