[IC‐P‐076]: MULTI‐CLASS DIFFERENTIAL DIAGNOSIS AMONG ALZHEIMER's, PARKINSON's, CORTICOBASAL SYNDROME AND PROGRESSIVE SUPRANUCLEAR PALSY
Bibliographic record
Abstract
The majority of previous computer-aided diagnosis (CAD) research focused on binary classification to discriminate a single neurodegenerative disorder (NDD) from healthy controls (CN). Given overlap among NDDs (e.g. AD, Frontotemporal dementia, Dementia with Lewy bodies, Corticobasal syndrome(CBS), Parkinson's (PARK) and Progressive Supranuclear Palsy (PSP)) in symptoms and atrophy, binary classification can lead to costly misdiagnosis. We have published early attempts on multi-class differential diagnosis among AD, FTD and CN [2]. In this study, we extend it to study 4 different NDDs (AD, PARK, CBS and PSP) from 3 datasets: Alzheimer disease neuroimaging initiative (ADNI) [5], the Parkinson's Progression Markers Initiative (PPMI) [6] and the 4 Repeat Tauopathy Neuroimaging Initiative (4RTNI) [7]. Baseline structural T1 MRI scans have been processed through Freesurfer (FS) pipeline. The cortical reconstructions were visually quality controlled to exclude erroneous segmentations. This provided 604 subjects (188 AD from ADNI, 39 PSP and 33 CBS from 4RTNI along with 243 PARK and 101 CN from PPMI), whose cortical thickness (CT) features were smoothed (FWHM=10mm) and resampled onto the fsaverage template. The 68 FS cortical labels were adaptively subdivided into smaller patches of approx. equal size (controlled by number of vertices) [3,4]. Patch-wise median CT served as input to the multi-class random forest (RF) classifier. Accuracy is evaluated using repeated holdout cross-validation (80% training) stratifying the training set [RHsT, 3] for different patch sizes (m=500, 1000, 2000, 3000, 5000 and 10000). The distribution of balanced accuracies for different m (200 trials of RHsT) are shown in Fig. 1. A non-parametric Wilcoxon Ranksum test comparing their performance revealed 1) they are significantly better than a chance classifier (20% accuracy in a 5-class experiment) and 2) they are not statistically significantly different from each other. The confusion matrix is shown in Fig. 2. The most discriminative patches for the RF classifier for m=1000 are shown in Fig. 3. Patch-wise CT from a single baseline MRI scan demonstrates differential diagnostic utility (balanced accuracy of 45%) in a challenging 5-class task among AD, PARK, CSB, PSP and CN. Distribution of the classification performance of the random forest classifier for different patch sizes m. The baseline performance of a random chance classifier (20% in a 5-class experiment) is indicated with a dashed line. A non-parametric Wilcoxon Ranksum test comparing their performance revealed 1) they are significantly better than a chance classifier and 2) they are not statistically significantly different from each other. The confusion matrix of the random forest classifier (derived from the predictions on the test set) for patch size m=1000, illustrates the challenge involved in a 5-class differential diagnosis task. Panel A on the left presents the confusion matrix in percentage (averaged from the 200 repetitions of RHsT), and panel B on the right presents raw average number (from 200 trials of RHsT) of test subjects in each class. We notice significant confusion among different NDDs, esp. between ALZ and PARK as well as PSP and CBS. Visualization of the importance (measured by their impact in the accuracy of RF classifier if removed from training) for cortical thickness features from different patches. Only areas with impact greater than 10% are highlighted, with their actual values displayed according to the colorbar. Darker pink patches are more discriminative than the lighter ones. This visualization reveals expected areas such as superior temporal, superior parietal, pericalcarine and insula (affected in AD, CBS) as well as postcentral (affected in Parkinsonian syndromes).
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.007 |
| Meta-epidemiology (narrow) | 0.002 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.003 | 0.003 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.003 | 0.002 |
| Research integrity | 0.003 | 0.001 |
| Insufficient payload (model declined to judge) | 0.017 | 0.009 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".