Uncertainty-Gated SHAP-Dirichlet Refinement for MRI-Based Depression Classification

Phuc Truong Loc Nguyen; Nguyen Truong Toan To; Truong Thanh Hung Nguyen; Hoang-Loc Cao; Hung Cao

eess.IVOct 19, 2026

Uncertainty-Gated SHAP-Dirichlet Refinement for MRI-Based Depression Classification

Authors: Phuc Truong Loc Nguyen, Nguyen Truong Toan To, Truong Thanh Hung Nguyen, Hoang-Loc Cao, Hung Cao

Organizations: FAU Erlangen-Nürnberg, Erlangen, Germany · University of New Brunswick, Fredericton, Canada

Abstract

Structural MRI-based classification of major depressive disorder (MDD) remains challenging due to subtle, heterogeneous anatomical changes. We propose uncertainty-gated SHAP-Dirichlet refinement, a post-hoc reliability layer that improves predictions without retraining the classifier. The method retrieves similar labeled SHAP explanations, constructs Dirichlet class evidence, and selectively refines uncertain predictions through probabilistic fusion, while allowing unreliable cases to be deferred for expert review. We evaluate five classifiers on 504 SRPBS subjects (252 controls and 252 MDD patients) using 159 anatomical features and repeated nested cross-validation. Refinement improved accuracy across all 8 repetitions for TabM and ResNet, yielding mean gains of 1.04 and 0.45 percentage points (sign test p=0.008), while modifying at most 8.3% of predictions. Gains were concentrated among low-confidence cases, with significant improvements in calibration. The method also benefited an RBF support vector machine but not logistic regression or random forests. Ablations confirmed the importance of uncertainty gating and soft fusion. Dirichlet evidence identified baseline errors (AUROC 0.62), while deferring 50% of cases increased retained-case accuracy to approximately 75%. These findings support selective explanation-guided refinement as a practical approach to improving prediction reliability and facilitating expert triage in MDD classification.

Figures & tables

Figure 1. Uncertainty-gated SHAP-Dirichlet refinement. (A) Structural MRI scans are processed into morphometric features, and a fixed classifier produces the baseline probability and label. (B) The query SHAP explanation is compared with a reference SHAP bank to retrieve explanation-similar subjects and estimate Dirichlet class evidence. (C) Local explanation stability decides whether the prediction is retained or refined, and unreliable subjects can be routed to expert review.

Classifier	Reps	Model	Accuracy	BAcc	MF1	AUC	Brier $\downarrow$	ECE $\downarrow$	I/R	$p$
TabM	4	Baseline	64.24	64.24	64.19	70.92	0.2217	7.56
		Layer, val. bank	64.29 (+0.05)	64.29 (+0.05)	64.25 (+0.05)	70.84 ( $-$ 0.08)	0.2196	7.52	2/4	0.95
		Layer, train+val. bank	65.28 (+1.04)	65.28 (+1.04)	65.25 (+1.06)	71.25 (+0.33)	0.2181	6.24	4/4	0.38
ResNet	4	Baseline	63.74	63.74	63.71	70.00	0.2338	11.44
		Layer, val. bank	64.24 (+0.50)	64.24 (+0.50)	64.22 (+0.51)	69.24 ( $-$ 0.75)	0.2300	8.11	3/4	0.65
		Layer, train+val. bank	64.19 (+0.45)	64.19 (+0.45)	64.18 (+0.47)	69.91 ( $-$ 0.09)	0.2287	9.03	4/4	0.68

Table 1. Out-of-fold test performance of the fixed baselines and the refinement layer with the pre-specified validation reference bank and with the enlarged training+validation bank, averaged over cross-validation repetitions (each pooled over

n=504

subjects). Accuracy, balanced accuracy (BAcc), macro F1 (MF1), and ROC-AUC are in %, with the change in percentage points in parentheses. ECE is in %. Green ( red ): better (worse) than the baseline of the same classifier. Reps: number of repetitions. I/R: repetitions with improved accuracy.

p

: corrected resampled

t

-test over all fold-level accuracy changes.

		TabM		ResNet
Component	Variant	$\Delta$	I/R	$\Delta$	I/R
Correction	Hard flip	$-$ 0.45	1/4	$-$ 0.35	0/4
	Hard or soft (validated)	+0.45	2/4	+0.40	3/4
	Soft fusion (ours)	+0.05	2/4	+0.50	3/4
Gate	None	$-$ 0.74	1/4	$-$ 0.20	1/4
	Random	$-$ 0.13	1/4	+0.11	1/4

Table 2. Ablations for the neural backbones. Each row changes one component of the pre-specified layer (soft fusion, validated RIS/ROS gate, validation bank,

K=5

), and all other choices are still selected on validation data.

\Delta

: mean accuracy change (percentage points) over repetitions, green for a gain and red for a loss. I/R: repetitions with improved accuracy. Shaded: the pre-specified layer.

Classifier	$\Delta$ Brier	$\Delta$ ECE	$\Delta$ NLL
TabM	$-$ 3.6 ∗ [ $-$ 6.2, $-$ 1.1]	$-$ 1.3 [ $-$ 2.5, 0.2]	$-$ 10.4 ∗ [ $-$ 18.2, $-$ 3.5]
ResNet	$-$ 5.1 ∗ [ $-$ 8.5, $-$ 1.9]	$-$ 2.4 ∗ [ $-$ 3.2, $-$ 0.7]	$-$ 16.2 ∗ [ $-$ 25.8, $-$ 7.4]
Logistic regression	$-$ 2.6 ∗ [ $-$ 5.0, $-$ 0.2]	$-$ 1.1 [ $-$ 3.1, 1.4]	$-$ 8.0 [ $-$ 19.0, 0.3]
RBF-SVM	$-$ 0.2 [ $-$ 2.6, 2.4]	+0.0 [ $-$ 2.2, 1.8]	+0.1 [ $-$ 5.7, 5.8]
Random forest	+0.9 [ $-$ 0.4, 2.3]	+0.2 [ $-$ 1.6, 1.9]	+1.9 [ $-$ 1.0, 4.9]

Table 3. Change in calibration after refinement with the enlarged bank (refined minus baseline, lower is better), averaged over repetitions, with 95% paired bootstrap intervals. Brier and NLL are

\times 10^{3}

, ECE is in percentage points. Green with ∗ : improvement whose interval excludes zero.

Figure 5. Reliability diagrams of the baseline and refined MDD probabilities (enlarged bank), pooled over repetitions. The dotted line marks perfect calibration.

		TabM			ResNet
Space	Metric	Prec.	Acc.	AUC	Prec.	Acc.	AUC
Raw features	Cosine	57.6	61.3	0.657	57.6	61.3	0.657
Raw features	Pearson	57.0	61.1	0.657	57.0	61.1	0.657
Raw features	Euclidean	56.2	60.2	0.650	56.2	60.2	0.650
SHAP	Cosine	58.1	62.3	0.668	58.7	62.5	0.674
SHAP	Pearson	57.4	61.2	0.657	57.9	62.2	0.662

Table 4. Quality of the retrieved evidence for test subjects with a validation reference bank and

K=5

. Prec.: percentage of retrieved references with the same true label as the query. Acc. and AUC: accuracy and ROC-AUC of

p^{\mathrm{Dir}}_{1}

as a stand-alone classifier. Shaded: SHAP space. Bold: best value in each column.

Score	TabM	ResNet
Low confidence	0.648 (0.62–0.67)	0.641 (0.62–0.67)
MC-dropout variance	0.601 (0.58–0.63)	0.628 (0.60–0.65)
RIS	0.510 (0.48–0.54)	0.522 (0.50–0.55)
ROS	0.532 (0.51–0.56)	0.496 (0.47–0.52)
SHAP-Dirichlet (Pearson)	0.613 (0.59–0.64)	0.618 (0.59–0.64)
SHAP-Dirichlet (Euclidean)	0.612 (0.59–0.64)	0.607 (0.58–0.63)

Table 5. AUROC for separating incorrect from correct baseline decisions on out-of-fold test predictions of all repetitions, with 95% bootstrap intervals. Shaded: scores that use SHAP-Dirichlet evidence. Bold: best value in each column.

Figure 6. Accuracy of the automatically handled test subjects when the least reliable subjects are deferred to expert review, with validation-percentile thresholds. Dotted line: accuracy without review.

Figure 7. A patient with MDD whose TabM baseline prediction was wrong and was corrected by the layer with the enlarged bank. Left: the seven largest SHAP values of the query (orange pushes toward MDD, blue toward HC). Right: weights of the five retrieved references and their true labels.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Classifier	Fixed settings and searched values
TabM	PLE (20 bins), mini ensemble of 32 members, averaged ensemble outputs, no normalization. Layer sizes [128] or [256], dropout 0.1/0.2/0.3, learning rate $5{\times}10^{-5}$ / $10^{-4}$ / $5{\times}10^{-4}$ / $10^{-3}$ , weight decay $10^{-4}$ / $5{\times}10^{-4}$ / $10^{-3}$ (72 configurations).
ResNet	PLE (20 bins), layer normalization, skip connections. Two blocks [128, 64, 32] or three blocks [256, 128, 64], dropout 0.1/0.2/0.3, learning rate as for TabM, weight decay $10^{-5}$ / $10^{-4}$ / $10^{-3}$ (72 configurations).
Logistic regression	Standardized features, L2 penalty, $C\in\{10^{-3},10^{-2},10^{-1},1,10\}$ .
RBF-SVM	Standardized features, Platt-scaled probabilities, $C\in\{0.1,1,10\}$ , $\gamma\in\{\text{scale},10^{-3},10^{-2}\}$ .
Random forest	300 trees, maximum depth $\{\text{none},5,10\}$ , minimum leaf size $\{1,5\}$ , maximum features $\{\sqrt{d},0.3d\}$ .

Appendix

Table 6. Hyperparameter grids. All grids were searched with a 3-fold inner cross-validation inside every outer fold, optimizing the F1-score.

Classifier	Folds	Distinct	Mean $\pm$ SD	Range
TabM	20	18	64.24 $\pm$ 4.48	56.00–73.27
ResNet	20	18	63.74 $\pm$ 3.42	56.00–69.00
Logistic regression	10	3	64.28 $\pm$ 3.89	58.00–70.30
RBF-SVM	10	6	65.87 $\pm$ 2.44	61.39–69.31
Random forest	10	6	66.28 $\pm$ 3.08	61.39–70.00

Appendix

Table 7. Selected models over all outer folds and repetitions. Distinct: number of distinct selected configurations. Accuracy: baseline test accuracy (%) of single outer folds.

Figure 8. t-SNE of the 504 test subjects (first TabM repetition) in feature space and SHAP space, with the 5-fold cross-validated 5-NN accuracy of each space.

			Val. bank			Train+val. bank
Model	Seed	Base	$\Delta$	C/B	$p$	$\Delta$	C/B	$p$
TabM	42	65.87	+0.20	7/6	1.00	+0.60	4/1	0.38
TabM	43	63.69	+1.59	11/3	0.06	+0.60	11/8	0.65
TabM	44	62.70	$-$ 0.40	10/12	0.83	+1.98	26/16	0.16
TabM	45	64.68	$-$ 1.19	5/11	0.21	+0.99	13/8	0.38
ResNet	42	63.10	+0.40	14/12	0.85	+0.40	19/17	0.87

Appendix

Table 8. Per-repetition results for the validation (Val.) and training+validation (Train+val.) banks. Base: pooled baseline accuracy (%).

\Delta

: accuracy change (pp), green for a gain and red for a loss. C/B: corrected/broken decisions.

p

: exact McNemar test.

		LogReg		SVM		RF
Comp.	Variant	$\Delta$	I/R	$\Delta$	I/R	$\Delta$	I/R
Correction	Hard flip	$-$ 0.10	0/2	+0.30	1/2	$-$ 0.99	0/2
	Hard or soft (validated)	$-$ 0.10	1/2	+0.99	1/2	$-$ 1.29	0/2
	Soft fusion (ours)	$-$ 0.40	1/2	+1.09	1/2	$-$ 0.99	0/2
Gate	None	$-$ 0.89	1/2	+0.40	1/2	$-$ 1.29	0/2
	Random	+0.02	1/2	+0.05	1/2	$-$ 0.90	0/2

Appendix

Table 9. Ablations for the classical classifiers (two repetitions).

\Delta

: mean accuracy change (pp), green for a gain and red for a loss. I/R: repetitions with improved accuracy. Shaded: the pre-specified layer.

Depression Detection

Depression detection uses language, behavior, speech, and biological signals to identify depressive symptoms and estimate their severity.

See all 76 papers

M2LG-DG: A Multi-modal Local-Global Domain Generalization Framework for Cross-site Major Depressive Disorder Classification
Aug 31, 2026Muhammad Asif Hasan, Yanming Zhu, Xuefei Yin +1Depression Detection Brain Disorder Classification
School of Information and Communication Technology, Griffith University, Gold Coast, QLD, 4215, Australia
Comparing Post-Hoc Explainable AI Methods for Interpreting Black-Box EEG Models in Depression Detection
May 27, 2026Antonia Šarčević, Nikolina FridGradient-Based Attribution Electroencephalography
University of Zagreb Faculty of Electrical Engineering and Computing, Croatia
MERID: Multimodal Exploration via Recursive Self-Improvement Agents for Major Depression Analysis
Sep 28, 2026Lei Liu, Zhaokang Liang, Qingcheng Zeng +4Depression Detection Multimodal Classification
Yale University · Zhejiang University · Northwestern University +3

Aug 31, 2026J/K move · Enter open · S save

M2LG-DG: A Multi-modal Local-Global Domain Generalization Framework for Cross-site Major Depressive Disorder Classification

Muhammad Asif Hasan, Yanming Zhu, Xuefei Yin, Alan Wee-Chung Liew

School of Information and Communication Technology, Griffith University, Gold Coast, QLD, 4215, Australia

Depression Detection Brain Disorder Classification Cross-Domain Generalization Neuroimaging Multimodal Disentangled Representation Learning Multimodal Fusion

Open paper AskPDF arXiv

Classification models based on resting-state functional magnetic resonance imaging (rs-fMRI) often show lower performance at imaging sites not included during model development, which can limit their use in clinical settings. Domain generalization (DG) addresses this issue by learning representations from source sites that remain effective for unseen target sites. However, existing DG approaches for psychiatric disorder classification commonly rely on a single imaging modality and may not fully account for site-specific acquisition effects on the learned representation space. Subjects scanned at the same site share scanner hardware, acquisition settings, and preprocessing characteristics, which can cause representations to reflect acquisition conditions rather than diagnostic information. In this work, we present M2LG-DG, a source-only multimodal local-global framework for cross-site major depressive disorder (MDD) classification. The framework employs a dual-stream rs-fMRI encoder, where the global pathway models inter-regional dependencies through self-attention and the local pathway performs graph-constrained aggregation over functional connectivity-derived brain graphs. Imaging and non-imaging representations are decomposed into shared and private components and integrated through bidirectional cross-attention with a learned modality gate. A cross-site supervised contrastive objective forms positive pairs from same-class subjects acquired at different source sites, encouraging the fused representation to preserve diagnostic information across acquisition domains. On four held-out REST-meta-MDD sites, M2LG-DG achieves an AUC of 69.48% and exceeds the closest comparison method by 2.18 percentage points. Experiments on the Autism Brain Imaging Data Exchange (ABIDE) dataset further support its applicability to other psychiatric neuroimaging classification tasks.

Abstract

Depression Detection

Explore similar work

M2LG-DG: A Multi-modal Local-Global Domain Generalization Framework for Cross-site Major Depressive Disorder Classification

Comparing Post-Hoc Explainable AI Methods for Interpreting Black-Box EEG Models in Depression Detection

MERID: Multimodal Exploration via Recursive Self-Improvement Agents for Major Depression Analysis