Task-evoked eye movements as candidate digital biomarkers of depression severity in adolescents and young adults

Youxiang Sun; Yujie Nie; Zhuoru Wu; Mengyuan Yu; Wenkong Wang; Xiang Li; Yumei Wang; Xin Ma

eess.SPOct 13, 2026

Task-evoked eye movements as candidate digital biomarkers of depression severity in adolescents and young adults

Authors: Youxiang Sun, Yujie Nie, Zhuoru Wu, Mengyuan Yu, Wenkong Wang, Xiang Li, Yumei Wang, Xin Ma,

Organizations: School of Control Science and Engineering, Shandong University, Jinan, Shandong, China · Engineering Research Center of Intelligent Unmanned System, Ministry of Education, China · Department of Psychiatry, The Chinese University of Hong Kong, Shatin, Hong Kong SAR, China · Department of Psychology, Shandong Provincial Hospital Affiliated to Shandong First Medical University, Jinan, Shandong, China

Abstract

Objective digital measures of depression severity could complement conventional symptom assessments and facilitate more timely clinical evaluation. Here, we investigated whether task-evoked eye movements provide interpretable behavioral markers of depression severity in adolescents and young adults. In this single-center, cross-sectional study, 155 participants aged 14--20 years completed emotional-spatial free-viewing and memory-guided recall tasks and were stratified by 17-item Hamilton Depression Rating Scale scores. We examined 19 prespecified eye-movement features across six task-informed domains using covariate-adjusted ordinal-trend models. Greater depression severity was associated with lower gaze-allocation entropy (standardized $β=-0.599$ ), fewer inter-area switches ( $β=-0.526$ ), longer first target-entry latency ( $β=0.572$ ), and lower target dwell proportion ( $β=-0.678$ ; all false-discovery-rate-adjusted $q<0.001$ ). Emotion-specific attentional-bias measures were not associated with severity, whereas associations involving fixation dispersion weakened after adjustment for anxiety symptoms. In a complementary machine-learning analysis incorporating 44 candidate eye-movement features and demographic covariates, a three-stage cascade classifier achieved the highest four-class stratification performance (accuracy, 0.665; macro-F1, 0.670; area under the receiver operating characteristic curve, 0.869), outperforming single-stage and two-stage approaches. Together, these findings identify altered visual exploration and memory-guided gaze control as candidate digital biomarkers of depression severity, rather than generalized negative attentional bias. These results support further prospective evaluation of task-evoked eye tracking as a complementary approach for timely severity assessment and clinical decision support.

Figures & tables

	No depression	Mild	Moderate	Severe	Statistic	$p$
	( $n=32$ )	( $n=39$ )	( $n=49$ )	( $n=35$ )
Age, years	17.22 (2.56)	17.23 (2.24)	16.06 (2.67)	15.60 (2.63)	$H=10.84$	0.013
Male sex, no. (%)	14 (43.8)	16 (41.0)	17 (34.7)	10 (28.6)	$\chi^{2}=2.08$	0.557
Education, years	11.34 (2.35)	11.18 (2.04)	10.14 (2.37)	9.69 (2.27)	$H=12.94$	0.005
HAMD-17 score	4.28 (2.44)	12.51 (3.09)	19.16 (2.22)	27.06 (3.87)	$H=144.02$	$<0.001$
HAMA score	5.09 (5.01)	9.31 (5.05)	14.10 (7.23)	18.37 (8.11)	$H=53.47$	$<0.001$

Table 1: Demographic and clinical characteristics across HAMD-17 symptom-severity groups.

Fig. 1: Representative free-viewing eye-movement features across HAMD-17 symptom-severity groups. Violin plots with embedded boxplots show unadjusted participant-level distributions for selected features from visual scanning coverage, emotion-related attentional bias, and gaze transition dynamics. Displayed $p$ values were obtained from ordered-severity regression models adjusted for age, sex, and education, with HC3 standard errors. The displayed $q$ values were adjusted across the 19 prespecified primary features.

Feature	Domain	Standardized $\beta$ (95% CI)	$p$	$q_{\mathrm{FDR}}$
Time to full AOI coverage	F1	0.305 (0.165, 0.445)	$<0.001$	$<0.001$
First-pass AOI coverage ratio	F1	$-0.447$ ( $-0.604$ , $-0.291$ )	$<0.001$	$<0.001$
AOI allocation entropy	F1	$-0.599$ ( $-0.733$ , $-0.466$ )	$<0.001$	$<0.001$
Negative–positive gaze allocation bias	F2	0.124 ( $-0.019$ , 0.266)	0.090	0.114
Negative AOI visit duration	F2	$-0.065$ ( $-0.238$ , 0.109)	0.466	0.492
Negative stickiness after first hit	F2	$-0.020$ ( $-0.184$ , 0.145)	0.813	0.813

Table 2: Associations between ordered HAMD-17 symptom severity and prespecified eye-movement features.

		Spearman correlation			Adjusted regression model
Feature	Domain	$\rho_{s}$	$p$	$q_{\mathrm{FDR}}$	Standardized $\beta$ (95% CI)	$p$	$q_{\mathrm{FDR}}$
Time to full AOI coverage	F1	0.278	$<0.001$	$<0.001$	0.299 (0.157, 0.441)	$<0.001$	$<0.001$
First-pass AOI coverage ratio	F1	$-0.405$	$<0.001$	$<0.001$	$-0.435$ ( $-0.586$ , $-0.283$ )	$<0.001$	$<0.001$
AOI allocation entropy	F1	$-0.651$	$<0.001$	$<0.001$	$-0.549$ ( $-0.682$ , $-0.416$ )	$<0.001$	$<0.001$
Negative–positive gaze allocation bias	F2	0.114	0.158	0.206	0.152 (0.011, 0.293)	0.035	0.049
Negative AOI visit duration	F2	$-0.095$	0.242	0.273	0.033 ( $-0.163$ , 0.230)	0.740	0.776

Table 3: Associations between continuous HAMD-17 scores and prespecified eye-movement features.

Fig. 2: Representative memory-guided recall features across HAMD-17 symptom-severity groups. Violin plots with embedded boxplots show unadjusted participant-level distributions for selected features from target acquisition, target-area gaze maintenance, and fixation stability. Displayed $p$ values were obtained from ordered-severity regression models adjusted for age, sex, and education, with HC3 standard errors. The displayed $q$ values were adjusted across the 19 prespecified primary features.

Protocol	Model	Accuracy	Macro F1	AUC
Single-stage four-class	TabPFN	0.619 (0.062)	0.627 (0.064)	0.865 (0.027)
Single-stage four-class	Avg(R8)	0.613 (0.032)	0.618 (0.035)	0.843 (0.014)
Single-stage four-class	Avg(13)	0.613 (0.051)	0.614 (0.047)	0.838 (0.014)
Single-stage four-class	SVM	0.594 (0.029)	0.594 (0.029)	0.843 (0.019)
Single-stage four-class	AG:NN	0.600 (0.093)	0.605 (0.094)	0.837 (0.027)
Two-stage cascade 2+2	LR $\rightarrow$ TabPFN	0.626 (0.018)	0.631 (0.022)	0.859 (0.025)

Table 4: Top-5 machine-learning configurations per protocol by composite ACC/F1/AUC rank.

Fig. 3: Top-5 machine-learning configurations for the four classification protocols. a , single-stage four-class (21 models); b , two-stage cascade 2+2 (441 configurations); c , two-stage cascade 1+3 (441 configurations); d , three-stage cascade 2+2+2 (9261 configurations). For each protocol, configurations are ordered by the sum of their ranks across accuracy, macro F1, and AUC (top 5 shown, left to right); grouped bars show the three metric values of each configuration. Values are means over five folds of stratified cross-validation. AG, AutoGluon; NN, NeuralNetTorch; ETE, ExtraTreesEntr; ETG, ExtraTreesGini; RFE, RandomForestEntr; LR, logistic regression; AG:LR, AutoGluon LinearModel; RF, random forest; SVM, support vector machine (RBF); TabPFN, TabPFN transformer; Avg(13), simple-average ensemble of 13 base AutoGluon models; Avg(T7), ensemble of the seven tree models; Avg(R8), ensemble of one representative per model family.

Fig. 4: Emotional-spatial free-viewing and memory-guided recall paradigm. Each trial comprised two to four displays containing positive, neutral, and negative faces, followed by an emotional-category cue and a six-location recall probe. The cue referred to the final face display immediately preceding it.

Task	Domain	Task-related interpretation	Primary features
Free-viewing	F1 Visual scanning coverage	Completeness and balance of visual sampling across the three emotional face AOIs	Time to full AOI coverage; first-pass AOI coverage ratio; AOI allocation entropy
Free-viewing	F2 Emotion-related attentional bias	Valence-related allocation and maintenance, particularly negative maintenance and delayed disengagement	Negative–positive gaze allocation bias; negative-AOI visit duration; negative stickiness after first hit
Free-viewing	F3 Gaze transition dynamics	Spatial and temporal organization of gaze across and within emotional AOIs	Inter-AOI switch rate; AOI transition diversity; transition tortuosity; within-AOI gaze dispersion
Recall	R1 Memory-guided target acquisition	Efficiency and accuracy of orienting to the cued location using encoded spatial information	First target-entry latency; initial target-entry accuracy; first target-fixation latency
Recall	R2 Target-area gaze maintenance	Preferential and sustained gaze on the recalled target after acquisition	Target dwell proportion; target lock-on ratio; post-hit off-target ratio
Recall	R3 Fixation stability	Local spatial stability of gaze within the target area	Target-area fixation dispersion; within-target fixation step size; target-area stability ellipse area

Table 5: Prespecified eye-movement feature domains.

Fig. S1: Top-5 classifiers for the exploratory two-class task (no/mild versus moderate/severe strata). Models are ordered by the sum of their ranks across accuracy, macro F1, and AUC (top 5 shown, left to right); grouped bars show the three metric values of each model. Values are means over five folds of stratified cross-validation. Abbreviations as in Fig. 3 .

Domain	Feature	Operational calculation	Interpretation
F1	Complete AOI coverage proportion	For each free-viewing display, code whether gaze entered all three emotional face AOIs at least once; average this binary value across displays.	Completeness of emotional AOI sampling during encoding; higher values indicate more complete coverage.
F1	AOI allocation entropy	Compute Shannon entropy of dwell-time proportions across negative, neutral, and positive AOIs for each display; average across displays.	Balance of gaze allocation across emotional AOIs; higher values indicate more even sampling.
F1	AOI allocation imbalance	Compute the L1 distance between the observed dwell-time distribution across the three emotional AOIs and a uniform distribution of 1/3, 1/3, 1/3; average across displays.	Deviation from balanced AOI allocation; higher values indicate more uneven sampling.
F1	First-pass AOI coverage	During the first pass through AOIs before the first revisit, count the number of unique emotional AOIs visited and divide by three; average across displays.	Breadth of initial visual sampling; higher values indicate broader first-pass coverage.
F1	Time to full AOI coverage	Compute the latency from display onset until all three emotional AOIs have been entered at least once; average across displays with full coverage.	Efficiency of completing emotional AOI coverage; higher values indicate slower coverage.
F2	Negative-AOI disengagement latency	After the first entry into the negative AOI, compute the duration until gaze first leaves the negative AOI; average across displays with a negative-AOI hit.	Disengagement from negative information; higher values indicate slower disengagement from the negative AOI.

Table S1: Operational definitions and interpretations of extracted eye-movement features.

Domain	Supplementary feature	Standardized $\beta$ (95% CI)	$p$	$q_{\mathrm{FDR}}$
F1	Complete AOI coverage proportion	$-0.538\,(-0.679,-0.398)$	$<0.001$	$<0.001$
F1	AOI allocation imbalance	$0.603\,(0.477,0.729)$	$<0.001$	$<0.001$
F2	Negative-neutral gaze allocation bias	$0.118\,(-0.031,0.267)$	$0.120$	$0.157$
F2	Negative AOI revisit count	$-0.215\,(-0.368,-0.062)$	$0.006$	$0.009$
F2	Negative-AOI disengagement latency	$-0.072\,(-0.238,0.093)$	$0.392$	$0.466$
F2	Late-early change in negative-positive gaze allocation bias	$-0.129\,(-0.295,0.036)$	$0.126$	$0.157$

Table S2: Supplementary results from ordered-severity models.

Feature domain	Supplementary feature	Spearman correlation			Adjusted regression model
		$\rho_{s}$	$p$	$q_{\mathrm{FDR}}$	$\beta$ (95% CI)	$p$	$q_{\mathrm{FDR}}$
F1	Complete AOI coverage proportion	$-0.600$	$<0.001$	$<0.001$	$-0.490\,(-0.627,-0.354)$	$<0.001$	$<0.001$
F1	AOI allocation imbalance	$0.614$	$<0.001$	$<0.001$	$0.550\,(0.418,0.682)$	$<0.001$	$<0.001$
F2	Negative–neutral gaze allocation bias	$0.151$	$0.060$	$0.088$	$0.113\,(-0.034,0.261)$	$0.132$	$0.171$
F2	Negative AOI revisit count	$-0.229$	$0.004$	$0.007$	$-0.230\,(-0.375,-0.085)$	$0.002$	$0.003$
F2	Negative-AOI disengagement latency	$-0.107$	$0.187$	$0.222$	$0.037\,(-0.152,0.226)$	$0.701$	$0.752$

Table S3: Associations between continuous HAMD-17 scores and supplementary eye-movement features.

Features domain	Features	Unadjusted model		HAMA-adjusted model		Extreme-excluded
		$\beta$ (95% CI)	$q_{\mathrm{FDR}}$	$\beta$ (95% CI)	$q_{\mathrm{FDR}}$	$\beta$ (95% CI)	$q_{\mathrm{FDR}}$
F1	Time to full AOI coverage	$0.284\,(0.143,0.425)$	$<0.001$	$0.265\,(0.084,0.446)$	$0.007$	$0.302\,(0.159,0.444)$	$<0.001$
F1	First-pass AOI coverage ratio	$-0.446\,(-0.596,-0.296)$	$<0.001$	$-0.530\,(-0.744,-0.316)$	$<0.001$	$-0.442\,(-0.603,-0.281)$	$<0.001$
F1	AOI allocation entropy	$-0.600\,(-0.732,-0.468)$	$<0.001$	$-0.747\,(-0.930,-0.565)$	$<0.001$	$-0.618\,(-0.750,-0.486)$	$<0.001$
F2	Negative-positive gaze allocation bias	$0.142\,(-0.029,0.313)$	$0.130$	$0.140\,(-0.080,0.359)$	$0.269$	$0.071\,(-0.070,0.211)$	$0.341$
F2	Negative AOI visit duration	$-0.077\,(-0.237,0.083)$	$0.364$	$-0.165\,(-0.368,0.038)$	$0.163$	$-0.107\,(-0.273,0.059)$	$0.247$

Table S4: Sensitivity analyses for primary eye-movement features.

	No ( $n=32$ )	Mild ( $n=39$ )	Moderate ( $n=49$ )	Severe ( $n=35$ )	$H$	$p$
Correct recall trials out of 12, mean (SD)	11.09 (1.06)	8.95 (2.05)	6.98 (2.84)	5.03 (2.99)	71.97	$<0.001$

Table S5: Recall-task performance across depression severity groups.

Depression Severity Estimation

Depression severity estimation infers symptom levels from clinical interviews, self-reports, speech, text, video, and other behavioral signals.

See all 35 papers

Apr 29, 2026

Entropy-Dominated Temporal Vocal Dynamics as Digital Biomarkers for Depression Detection

Automated depression detection often relies on static aggregation of conversational signals, potentially obscuring clinically meaningful behavioral dynamics. We investigated whether entropy-driven temporal biomarkers improve depression detection beyond standard pooled features using the DAIC-WOZ corpus. Using 142 labeled participants, we reconstructed utterance-level acoustic trajectories and compared pooled temporal baselines, trajectory dynamics, Shannon entropy biomarkers, recurrence quantification, sample entropy, fractal complexity, and coupling biomarkers under leakage-aware validation. Static pooling achieved an AUC of 0.593, trajectory dynamics improved performance to 0.637, and entropy biomarkers produced the strongest statistically significant improvement over pooled baselines (AUC 0.646; nested cross-validated AUC 0.615; permutation p = 0.017). Entropy biomarkers outperformed recurrence, coupling, sample entropy, and fractalbased features, with several biomarkers stable across folds. These findings suggest depression-related signal may lie less in average acoustic levels than in entropy of conversational dynamics, supporting temporally informed digital phenotypes for mental-health assessment.

Himadri S Samanta

Independent Researcher, Austin, Texas, USA

Depression Detection Mental Health Clinical Prediction

Ask PDF

Jul 28, 2026

Depression Markers in Speech: An Approach based on Tract Variables Dynamics

This study identifies new depression biomarkers based on the dynamical properties of tract variables, which represent geometric features describing the configuration of the speech articulators. A key advantage of this approach lies in its ability to quantify aspects of the articulatory process that have not been previously explored in the context of depression, namely predictability, complexity, and randomness. These properties are respectively characterised using the Largest Lyapunov Exponent, the Correlation Dimension, and the Sample Entropy. Thorough experiments were conducted on the Androids Corpus, a publicly available dataset comprising 64 speakers diagnosed with depression by clinicians and 54 control speakers with no reported history of mental health conditions. The results indicate that the proposed biomarkers effectively discriminate between the depressed and control speakers, as evidenced by the high Cliffs delta values across both read and spontaneous speech.

Sahar Altalhi, Tanaya Guha, Alessandro Vinciarelli

University of Glasgow (United Kingdom) and Taif University (Saudi Arabia) · University of Glasgow (United Kingdom)

Depression Detection Speech Processing

Ask PDF

Apr 11, 2026

Context Matters: Vision-Based Depression Detection Comparing Classical and Deep Approaches

The classical approach to detecting depression from vision emphasizes interpretable features, such as facial expression, and classifiers such as the Support Vector Machine (SVM). With the advent of deep learning, there has been a shift in feature representations and classification approaches. Contemporary approaches use learnt features from general-purpose vision models such as VGGNet to train machine learning models. Little is known about how classical and deep approaches compare in depression detection with respect to accuracy, fairness, and generalizability, especially across contexts. To address these questions, we compared classical and deep approaches to the detection of depression in the visual modality in two different contexts: Mother-child interactions in the TPOT database and patient-clinician interviews in the Pitt database. In the former, depression was operationalized as a history of depression per the DSM and current or recent clinically significant symptoms. In the latter, all participants met initial criteria for depression per DSM, and depression was reassessed over the course of treatment. The classical approach included handcrafted features with SVM classifiers. Learnt features were turn-level embeddings from the FMAE-IAT that were combined with Multi-Layer Perceptron classifiers. The classical approach achieved higher accuracy in both contexts. It was also significantly fairer than the deep approach in the patient-clinician context. Cross-context generalizability was modest at best for both approaches, which suggests that depression may be context-specific.

Maneesh Bilalpur, Saurabh Hinduja, Sonish Sivarajkumar +4

Intelligent Systems Program, University of Pittsburgh, Pittsburgh, USA · CGI Technologies and Solutions Inc, USA · University of Oregon, USA +3

Depression Detection Domain Generalization Algorithmic Fairness

Ask PDF

Abstract

Depression Severity Estimation

Explore similar work

Entropy-Dominated Temporal Vocal Dynamics as Digital Biomarkers for Depression Detection

Depression Markers in Speech: An Approach based on Tract Variables Dynamics

Context Matters: Vision-Based Depression Detection Comparing Classical and Deep Approaches