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

Lei Liu; Zhaokang Liang; Qingcheng Zeng; Chenda Duan; Lu Mi; Zhen Tan; Tianyu Liu

cs.AISep 28, 2026

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

Authors: Lei Liu, Zhaokang Liang, Qingcheng Zeng, Chenda Duan, Lu Mi, Zhen Tan, Tianyu Liu

Organizations: Yale University · Zhejiang University · Northwestern University · University of California, Los Angeles · Tsinghua University · Stevens Institute of Technology

Abstract

Major depressive disorder (MDD) severely impacts daily activities and quality of life. Detecting MDD involves multimodal data, such as interview recordings and sensor measurements. This is particularly challenging, as these heterogeneous modalities often demand distinct, customized prediction pipelines. Existing efforts to address this challenge have explored both manually engineered multimodal architectures and agent-assisted pipeline development. Despite their progress, it remains challenging to autonomously revise pipelines based on experimental feedback and carry verified improvements forward into subsequent designs. To this end, we propose Multimodal Exploration via Recursive Self-Improvement Agents for Major Depression Analysis (MERID). The framework develops depression pipelines through experience-based recursive self-improvement (RSI). Grounded State Construction (GSC) grounds experience by aligning multimodal records with subject-level depression targets. Coupled Pipeline Exploration (CPE) jointly modifies representations, fusion, and predictors to build successor pipelines for classification and severity estimation. Evidence-Guided Evolution (EGE) guides revisions through feedback and verifies gains under uncertainty in small depression cohorts before inheritance. Extensive experiments on depression benchmarks show that MERID achieves the best results on multiple tasks compared with multimodal and agent-based baselines. Further analysis highlights the value of acoustic and linguistic cues for depression detection. Our code is available at https://github.com/DiscoAILab/MERID

Figures & tables

Figure 2: Conventional approaches require repeated experiments and manual pipeline tuning for depression detection. MERID uses experimental feedback to revise pipelines and verify improvements. Retained pipelines and experience guide recursive self-improvement for depression assessment.

Figure 3: The framework of the proposed MERID . GSC aligns experimental records, and CPE jointly explores pipeline components. EGE uses feedback to guide revisions and verifies gains under uncertainty. The retained pipeline and updated experience guide the next RSI round.

Dataset	E-DAIC		MPDD-2025		Mental Health	Kaggle
Method	Binary F1 $\uparrow$	PHQ-8 CCC $\uparrow$	Elder Track $\uparrow$	Young Track $\uparrow$	Status F1 $\uparrow$	A–V F1 $\uparrow$	Avg. rank $\downarrow$
MulT	0.5221	-0.0047	0.4549	0.3952	—	—	4.50 (4)
Flex-MoE	0.5993	0.0135	0.5440	0.4483	0.9548	0.6201	2.67 (6)
MCMoE	0.4105	0.0000	0.4511	0.3419	n/a	n/a	5.75 (4)
Gemini 2.5 Flash	0.3978	-0.0917	0.4322	0.3995	0.0843	0.2966	5.33 (6)
MDAgents	0.4697	0.0520	0.2096	0.4069	0.0607	0.3845	4.33 (6)

Table 1: Performance comparison on depression benchmarks and supplementary tasks, with five-seed means for MERID using Claude Opus 4.8 and GPT-6 Astra. Best and second-best results are bold and underlined . See Appendices B.1 and B.2 for per-seed results.

Figure 4: Five-seed score differences from the best baseline in Table 1 (markers: seeds; bars: means). Positive values favor MERID . The axis is linear within

\pm 0.1

and logarithmic beyond.

Panel of Figure 5 — Figure 5: Ablation visualization and sensitivity analyses with Opus 4.8; see Appendix B.5 .

Figure 7: RSI source evolution (a) and score–cost comparisons (b,c). MERID points are five-seed means. Details: Appendices C.2 and D .

Appendix figures & tables33 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: A recorded MPDD-2025 evidence-guided evolution step. Dialogue is condensed from the review record. The final arrow tracks the best development candidate, which is distinct from the retained incumbent.

Aspect	DepressionAgent (text branch)	MERID
Evidence	Interview statements and supporting or countervailing interpretations	Task constraints, candidate scores, prediction diagnostics, and execution checks
Revision target	The current participant’s risk judgment	Representation–fusion recipes and predictor configurations
Depression-specific action	Re-read the transcript for overlooked risk evidence	Check clinical-scale exclusion and propose an ordinal severity predictor
Verification	Reconsider evidence through review and re-evaluation	Evaluate proposed pipelines under a shared protocol and replacement margin
Retained consequence	A final case judgment and its reasoning record	A reusable pipeline, admitted components, and experience for the next round

Appendix

Table 2: How depression-related evidence enters revision. DepressionAgent describes the evaluated text-branch adaptation. The MERID column describes its pipeline-development loop, with the concrete example in Figure 8 .

Task	Metric	Seed 0	Seed 1	Seed 2	Seed 3	Seed 4	Mean $\pm$ SD
E-DAIC binary	Macro-F1	0.6655	0.6937	0.6937	0.6655	0.6655	0.6768 $\pm$ 0.0154
E-DAIC PHQ-8	CCC	0.3201	0.3201	0.3538	0.3201	0.3201	0.3268 $\pm$ 0.0151
MPDD-2025 Elder	Track score	0.5820	0.6071	0.6177	0.6282	0.6335	0.6137 $\pm$ 0.0204
MPDD-2025 Young	Track score	0.4116	0.3545	0.3610	0.3755	0.3594	0.3724 $\pm$ 0.0233
Mental Health	Macro-F1	0.9704	0.9704	0.9630	0.9527	0.9710	0.9655 $\pm$ 0.0079
Kaggle A–V	Macro-F1	0.7811	0.7624	0.7803	0.8074	0.7869	0.7836 $\pm$ 0.0162

Appendix

Table 3: Test scores of MERID under seeds 0–4 for every column of Table 1 . Each seed is one complete run (development search, review and a single test evaluation). Best and second-best distinct seed scores within each row are bold and underlined, including ties. The last column reports the mean and sample standard deviation over the five runs.

Figure 10: Test scores under seeds 0–4 with Claude Opus 4.8 (blue) and GPT-6 Astra (orange) as the base model; bars are seed means. Each panel keeps its task’s metric and scale.

Task	Metric	Opus 4.8	GPT-6 Astra	$\bm{\Delta}$
E-DAIC binary	Macro-F1	0.6768 $\pm$ 0.0154	0.7057 $\pm$ 0.0118	+0.0289
E-DAIC PHQ-8	CCC	0.3268 $\pm$ 0.0151	0.3201 $\pm$ 0.0000	-0.0067
MPDD-2025 Elder	Track score	0.6137 $\pm$ 0.0204	0.6148 $\pm$ 0.0291	+0.0011
MPDD-2025 Young	Track score	0.3724 $\pm$ 0.0233	0.3714 $\pm$ 0.0261	-0.0010
Mental Health	Macro-F1	0.9655 $\pm$ 0.0079	0.9565 $\pm$ 0.0042	-0.0090
Kaggle A–V	Macro-F1	0.7836 $\pm$ 0.0162	0.6070 $\pm$ 0.0000	-0.1766

Appendix

Table 4: Test scores of MERID with Claude Opus 4.8 and GPT-6 Astra as the base model: mean

\pm

sample SD over seeds 0–4 of the same 16 run units, protocol and code.

\Delta

is Astra minus Opus.

Quantity	Opus 4.8	GPT-6 Astra
Run attempts for 80 tasks	80	85
Aborted attempts	0	5
MPDD-2026 proposals audited	112	135
invalid (run aborts)	0	5 (3.7%)
duplicates (re-asked once)	0	0
Chair replies re-asked for format	35	0

Appendix

Table 5: Failure analysis of the two five-seed studies. An attempt is one run of a task; aborted attempts were re-run under the same seed, at most twice. MPDD-2026 audits every proposal: an invalid one ends the run, a duplicate is re-asked once. The chair is re-asked when its reply lacks the required format. Token counts and costs cover completed runs; costs use uncached list prices, an upper bound when part of the input was cached.

Task	Seed	Att.	Prop.	Failure	Offending fragment
MPDD-2026 Elder	0	1	7	out-of-space value	"sev_fusion": "wav2emb:densenet"
MPDD-2026 Elder	2	1	6	out-of-space value	"audio": "wav2span"
MPDD-2026 Elder	2	2	1	out-of-space value	"audio": "wav2.0"
MPDD-2026 Elder	4	1	6	malformed JSON	gp_stats_pz","model":"plsn4":"bad"}}
MPDD-2026 Young	2	1	9	malformed JSON	1.0","sev_model":"gb0.05d1,"sev_fusion":

Appendix

Table 6: The invalid GPT-6 Astra proposals that aborted a run: task, seed, attempt, the proposal’s position in the run, and the offending part of its final JSON line, checked against the action space.

Task / test metric $\uparrow$	w/o CPE	w/o EGE	MERID (full)
E-DAIC binary (F1)	0.5151 $\pm$ 0.0760	0.5162 $\pm$ 0.0345	0.6768 $\pm$ 0.0154
E-DAIC PHQ-8 (CCC)	0.0361 $\pm$ 0.0452	0.2997 $\pm$ 0.0196	0.3268 $\pm$ 0.0151
Mental Health (F1)	0.9672 $\pm$ 0.0149	0.9530 $\pm$ 0.0005	0.9655 $\pm$ 0.0079
MPDD-25 Elder 1 s binary	0.7231 $\pm$ 0.0846	0.7850 $\pm$ 0.0042	0.7593 $\pm$ 0.0349
MPDD-25 Elder 1 s ternary	0.6089 $\pm$ 0.0822	0.5955 $\pm$ 0.0667	0.6295 $\pm$ 0.0673
MPDD-25 Elder 1 s quinary	0.4583 $\pm$ 0.0341	0.4472 $\pm$ 0.0482	0.4738 $\pm$ 0.0647

Appendix

Table 7: Module ablations. Test results report mean

\pm

sample SD over seed means: MERID (full) is Table 1 ’s five-seed result (seeds 0–4); w/o CPE and w/o EGE use seeds 0–2 and start without an inherited pipeline. Repeated runs are averaged within each seed. MPDD-2025 uses the official composite score. GSC results use development data with one run per configuration. Best and second-best distinct means within each row are bold and underlined .

Binary classification
Available	Dev Macro-F1	Test Macro-F1 $\uparrow$	Test Accuracy $\uparrow$	Deployed
A	0.6348 $\pm$ 0.0000	0.5336 $\pm$ 0.0679	0.5655 $\pm$ 0.0412	A
V	0.5754 $\pm$ 0.0476	0.6086 $\pm$ 0.0092	0.6488 $\pm$ 0.0103	V
T	0.6750 $\pm$ 0.0531	0.6107 $\pm$ 0.0535	0.6607 $\pm$ 0.0309	T
AV	0.6348 $\pm$ 0.0000	0.5612 $\pm$ 0.1064	0.6726 $\pm$ 0.0103	A+V
AT	0.7348 $\pm$ 0.0000	0.6468 $\pm$ 0.0153	0.6964 $\pm$ 0.0179	A+T

Appendix

Table 8: Available modalities on E-DAIC. Values are mean

\pm

sample SD over three seeds. Development scores describe the selected candidate; test scores describe the locked deployment. PHQ-8 development RMSE is the negated selection score. Deployed sources are the union across ensemble members within a run; semicolons separate configurations observed across seeds. Best and second-best test means within each task and metric are bold and underlined.

Figure 11: Additional sensitivity scans from the recorded configurations. Top: E-DAIC PHQ CCC versus budget, maximum revision rounds, and seed. Middle: MPDD-2025 1 s ternary official scores versus folds, budget, and maximum rounds. Bottom: MPDD-2026 joint track scores versus budget, maximum rounds, and seed. Points average available runs at each setting, and whiskers show repeated-run ranges where nonzero. Panels use separate metric scales. These are separate runs at different settings, rather than successive states of one trajectory.

Selection rule	Development	Test
Incumbent $+$ one standard error	0.7172	$\mathbf{0.6843}\;[0.6655,0.6937]$
Classic one-standard-error rule	0.7737	0.4991
Maximum development score	0.7737	0.4876

Appendix

Table 9: Selection-rule sensitivity on E-DAIC binary classification. The incumbent-based rule reports the mean and range of three test runs; alternative rules were each evaluated once. Development and test scores are Macro-F1 (

\uparrow

Task / metric	Reference $+$ one SE	Winner’s-curse configuration	Search selection (one SE)
E-DAIC binary (Macro-F1)	0.6937	0.6814	0.5492
E-DAIC PHQ-8 (CCC)	0.3201	0.3201	0.3201
Mental Health (Macro-F1)	0.9630	0.9535	0.9535
MPDD-25 Elder 1 s binary	0.7212	0.7212	0.7212
MPDD-25 Elder 1 s ternary	0.6842	0.6842	0.6220
MPDD-25 Elder 1 s quinary	0.3962	0.5490	0.4011

Appendix

Table 10: Historical deployment configurations, with one seed-0 run per cell. MPDD-2025 uses the official composite. The Mental Health winner’s-curse-family record predates that gate and uses only the rank-transfer check. Best and second-best distinct values per row are bold and underlined. Each cell uses the latest logged run for its protocol, configuration, task, and seed.

Figure 12: Selection and verification analyses. (a,b) E-DAIC selection-rule and revision records; the test whisker spans three incumbent-based runs, while each alternative has one run. (c,d) A separate terminal-margin replay of five proposed promotions. E-DAIC uses Macro-F1; MPDD-2025 Elder uses its official score. Each task is shown separately. Panel (d) compares saved promoted single-model predictions with the recorded reverted deployment.

Task	Full	Staged	Fixed recipe	Fixed predictor	1-SE	Random predictor	Random recipe
E-DAIC binary	0.6937	0.4991	0.6814	0.4593	0.4991	0.5564	0.5735
E-DAIC PHQ-8	0.3201	0.3201	0.3201	0.3201	0.3201	0.3307	0.3201
Mental Health	0.9630	0.9711	0.9527	0.9305	0.9535	0.9527	0.9539
MPDD-25 binary	0.7212	0.7543	0.7212	0.7212	0.7565	0.7266	—
MPDD-25 ternary	0.6390	0.6443	0.6922	0.6482	0.6183	0.6619	—
MPDD-25 quinary	0.4115	0.4066	0.4514	0.4317	0.4169	0.4624	—

Appendix

Table 11: Search and selection variants. Ordinary arms retain one seed-0 run per task; random-space arms average three seeds. MPDD-2025 rows then average the 1 s and 5 s Elder tasks. Metrics are Macro-F1 for binary/status classification, CCC for PHQ-8, and the official composite for MPDD-2025. MPDD random recipes are inapplicable because the proposed recipe space already exhausts the 12 eligible combinations. Best and second-best distinct values are bold and underlined. For each protocol, configuration, task, and seed, the latest logged run is used before seed and window aggregation.

Task	Memory off	Evolving	$\Delta$	Calls (off $\to$ evolving)
E-DAIC binary (F1)	0.4927	0.5099	+0.0172	19.7 $\to$ 33.7
E-DAIC PHQ-8 (CCC)	0.3142	0.3311	+0.0168	28.7 $\to$ 29.3
Mental Health (F1)	0.9647	0.9582	-0.0065	27.0 $\to$ 33.0
MPDD-25 Elder 1 s binary	0.7802	0.7886	+0.0084	22.0 $\to$ 18.0
MPDD-25 Elder 1 s ternary	0.6557	0.6335	-0.0222	18.0 $\to$ 39.7
MPDD-25 Elder 1 s quinary	0.4164	0.4164	+0.0000	22.0 $\to$ 25.7

Appendix

Table 12: Research-memory ablation. Test scores and LLM calls are means over three seeds.

\Delta

is evolving minus off on the indicated task metric.

Figure 13: Retained development scores across recorded revision steps in the separate research-memory experiment. Colors distinguish memory settings, and line styles and markers identify seeds. Each panel includes all six task-specific runs. Scores are Macro-F1 for E-DAIC binary and Mental Health, CCC for E-DAIC PHQ, and the official composite for MPDD-2025 Elder. Curves stop at their last record, and overlapping trajectories retain their measured values.

Memory	Step	Runs	New evaluations	Retained-score gains
Off	1	23	751	2
Off	2	5	107	0
Off	3	3	52	0
Evolving	1	22	727	3
Evolving	2	12	270	0
Evolving	3	8	157	0

Appendix

Table 13: Candidate evaluations and retained-score gains by recorded revision step. Each run contributes only to steps present in its summary. The step counts cover the 54-run research-memory experiment.

Figure 14: Source and memory measures in the amended eight-round study. (a-c) Six E-DAIC runs per arm; curves average available records with

\pm

1 standard-error bands. Panel (a) is candidate-weighted L2; it differs from the source-slot proportions in the main figure. Gray lines in (b,c) are binary and PHQ-8 training-label references. (d) Lessons and review minutes from nine runs per review arm. (e) Citation rates use nine memory-enabled runs; action matching uses the six E-DAIC runs with applicable entries. (f) Run counts for panel (a), with rows following the arm legend. Counts describe available task-run records, not additional independent seeds.

Figure 15: Development and test trajectories under the amended eight-round protocol. Rows show E-DAIC binary classification, E-DAIC PHQ-8, and MPDD-2025 Elder 1 s ternary classification. Left: best newly evaluated candidate (solid) and retained incumbent (dashed), using available development records. Right: the locked deployment at each round. After a run ends, its final deployment score is carried forward through round eight; unrecorded intermediate rounds remain excluded from that round’s mean. Each task and arm has three runs; intermediate test means use one to three available records. Bands show

\pm

1 standard error when at least two records are available. Each task retains its own metric scale.

Task / test metric	Evolving $-$ random	Evolving $-$ no memory
E-DAIC binary / Macro-F1	$+0.0004$ [-0.0176, 0.0185]	$-0.0017$ [-0.0206, 0.0173]
E-DAIC PHQ-8 / CCC	$+0.0024$ [-0.0023, 0.0071]	$+0.0045$ [0.0013, 0.0079]
MPDD-25 Elder 1 s ternary / official	$+0.0076$ [-0.0093, 0.0245]	$+0.0002$ [-0.0121, 0.0125]

Appendix

Table 14: Secondary L4 slope contrasts per round. Each arm has three runs per task. Brackets are unadjusted 95% run-bootstrap intervals; task metrics are kept separate.

Task	Seed	Add / revise / retire	Citation rate	Action-matched use
E-DAIC binary	0	14 / 12 / 2	0.73	1 / 7
E-DAIC binary	1	11 / 15 / 2	0.82	6 / 22
E-DAIC binary	2	17 / 10 / 1	0.72	6 / 21
E-DAIC PHQ-8	0	14 / 15 / 0	0.80	1 / 5
E-DAIC PHQ-8	1	14 / 12 / 1	0.66	7 / 33
E-DAIC PHQ-8	2	8 / 6 / 2	0.81	6 / 14

Appendix

Table 15: Research-memory use in the nine evolving eight-round runs. Operations count entries added, revised and retired over the run. Citation rate is the fraction of active entries whose identifier appears in the next round’s minutes or proposals. Action matching counts activated entries whose next action names a source family and whose next round adds that family, over the activated entries that name one; this source-family proxy is computed for E-DAIC only.

Task	Arm	Seed	Reviews	Failed	Requested / declined	New rounds	Candidates
E-DAIC binary	Memory	0	8	1	3 / 1	6	448
E-DAIC binary	Memory	1	8	1	1 / 0	7	512
E-DAIC binary	Memory	2	8	1	2 / 1	6	512
E-DAIC binary	No memory	0	8	0	6 / 0	8	512
E-DAIC binary	No memory	1	8	1	4 / 0	7	480
E-DAIC binary	No memory	2	8	1	5 / 0	7	480

Appendix

Table 16: Review and exploration counts for every amended eight-round run. Reviews counts scheduled review rounds, including failures. Requested / declined counts explicit follow-up requests for a revision and requests that still produced none. New rounds counts revision rounds with newly evaluated candidates. Random revision exhausts its drawable MPDD-2025 space after five rounds.

Figure 16: LLM efficiency. (a) E-DAIC test Macro-F1 versus recorded calls: MERID uses the means over its five seed runs; inference baselines cover 56 test participants. DepressionAgent uses its scored 268-call run. Non-LLM methods appear at zero calls. (b,c) Diamonds show medians of nine task-level means from the module ablations, with full MERID from the five-seed runs of Table 1 ; vertical lines show their minimum-to-maximum ranges, not uncertainty intervals. Local fitting and feature extraction are outside this accounting.

Task	MERID (full)	w/o CPE	w/o EGE	Runs
E-DAIC binary	11.4 / $0.60	83.0 / $4.67	2.0 / $0.16	5 / 1 / 3
E-DAIC PHQ-8	11.4 / $0.61	41.0 / $2.12	1.0 / $0.06	5 / 1 / 3
Mental Health	10.8 / $0.64	89.0 / $5.09	0.0 / $0.00	5 / 1 / 3
MPDD-25 Elder 1 s binary	12.2 / $0.65	89.0 / $5.29	1.0 / $0.01	5 / 1 / 1
MPDD-25 Elder 1 s ternary	12.4 / $0.70	20.0 / $1.06	1.0 / $0.01	5 / 1 / 1
MPDD-25 Elder 1 s quinary	12.2 / $0.69	33.0 / $1.89	1.0 / $0.01	5 / 1 / 1

Appendix

Table 17: LLM usage per run of the module ablations: mean LLM calls / USD over the logged runs of each task and configuration (Claude Opus 4.8 at

5/

25 per million input/output tokens). Local feature extraction and model fitting are excluded. Full MERID uses the five runs of Table 1 ; for the removals, recovered runs without a cost record are not included, so their sample differs from the performance sample of Table 7 . Runs gives the number of logged runs per configuration (full / w/o CPE / w/o EGE).

Memory	Runs	Cap	Total mean	Total range	Revision mean
Memory off	27	320	264.4	139–320	33.7
Evolving memory	27	320	272.9	139–320	42.7

Appendix

Table 18: Realized candidate evaluations in the 54-run research-memory comparison. Each arm covers nine tasks and three seeds. Means are over runs; ranges span their totals. Revision evaluations exclude initialization.

Task	Arm	Cap	Initial mean	Total mean	Total range
E-DAIC binary	Memory	512	254.7	490.7	448–512
E-DAIC binary	No memory	512	256.0	490.7	480–512
E-DAIC binary	Random	512	256.0	512.0	512–512
E-DAIC PHQ-8	Memory	512	177.3	512.0	512–512
E-DAIC PHQ-8	No memory	512	182.0	512.0	512–512
E-DAIC PHQ-8	Random	512	186.7	512.0	512–512

Appendix

Table 19: Exploration effort in the amended eight-round study. Each row covers three seeds. The cap includes initialization and revisions. Counts include all recorded candidate evaluations. The total range is the observed minimum–maximum across the three runs.

Dataset	Cohort	Task	Metric	Official	MulT	Flex-MoE	Gemini 2.5 Flash	MDAgents	Ours
E-DAIC	–	Binary	Accuracy $\uparrow$		0.5893	0.7143	0.6607	0.6429	0.7393
			Bal. Acc. $\uparrow$		0.5226	0.5958	0.4744	0.4947	0.6702
			Macro-F1 $\uparrow$		0.5221	0.5993	0.3978	0.4697	0.6768
			W. F1 $\uparrow$		0.5925	0.6836	0.5541	0.5887	0.7326
			$\kappa$ $\uparrow$		0.0445	0.2209	-0.0683	-0.0127	0.3554
		PHQ-8	RMSE $\downarrow$		6.9902	7.0394	7.7113	6.6201	5.9098

Appendix

Table 20: Complete performance metrics, continued over three parts. MERID reports means over seeds 0–4 (Kaggle Video: retained run); baselines are unchanged. This part reports E-DAIC, MPDD-2026, Mental Health, and Kaggle. Arrows indicate preferred directions. Best and second-best distinct values among the methods shown are bold and underlined, including ties; dashes denote unavailable results. Bal. Acc. and W. F1 denote balanced accuracy and weighted F1. Avg. Rank uses one primary metric per task: Macro-F1 for classification, RMSE for PHQ regression, and the official score for MPDD-2025. Per-task ranks use midranks among reported methods; averages require complete task coverage and exclude track aggregates. MPDD-2026 denotes MPDD-AVG 2026. Its MulT and Flex-MoE PHQ-9 estimates are derived from ternary-class midpoints. Kaggle evaluates acted-emotion proxy labels; MERID uses actor-disjoint nested cross-validation with review disabled.

Dataset	Cohort	Task	Metric	Official	MulT	Flex-MoE	Gemini 2.5 Flash	MDAgents	Ours
MPDD-2025	Elder	1 s Binary	Accuracy $\uparrow$	0.6916	0.7841	0.7416	0.7357	0.1982	0.7789
			Macro-F1 $\uparrow$	0.5900	0.6318	0.6266	0.5048	0.1809	0.6964
			W. Acc. $\uparrow$	0.6360	0.6332	0.6597	0.5055	0.5035	—
			W. F1 $\uparrow$	0.7222	0.7852	0.7604	0.7238	0.1038	0.7987
			Task Acc. $\uparrow$	0.6638	0.7086	0.7006	0.6206	0.3509	0.7712
			Task F1 $\uparrow$	0.6561	0.7085	0.6935	0.6143	0.1424	0.7475

Appendix

Table 21: Complete performance metrics (continued): MPDD-2025 Elder. All seven task metrics and all three track aggregates are shown. W. Acc. denotes inverse-class-frequency-weighted accuracy, and W. F1 denotes support-weighted F1. Task Acc. averages Accuracy and W. Acc.; Task F1 averages Macro-F1 and W. F1; Official averages Task Acc. and Task F1. Track metrics average the corresponding task metrics over both windows. Flex-MoE cells report the means of three runs. Ours Track Acc. and Track F1 are computed from the reported task components. Ours W. Acc. was not included in the available main-run summary and is left unreported. Formatting follows the first part.

Dataset	Cohort	Task	Metric	Official	MulT	Flex-MoE	Gemini 2.5 Flash	MDAgents	Ours
MPDD-2025	Young	1 s Binary	Accuracy $\uparrow$	0.5644	0.4924	0.6048	0.4848	0.5455	0.4038
			Macro-F1 $\uparrow$	0.5446	0.4491	0.5749	0.3450	0.4762	0.4018
			W. Acc. $\uparrow$	0.5644	0.4924	0.6048	0.4848	0.5455	—
			W. F1 $\uparrow$	0.5446	0.4491	0.5749	0.3450	0.4762	0.4018
			Task Acc. $\uparrow$	0.5644	0.4924	0.6048	0.4848	0.5455	0.4038
			Task F1 $\uparrow$	0.5446	0.4491	0.5749	0.3450	0.4762	0.4018

Appendix

Table 22: Complete performance metrics (continued): MPDD-2025 Young. Metric definitions, aggregation, and formatting follow the Elder part. Flex-MoE cells report the means of three runs. Avg. Rank covers all ten MPDD-2025 tasks across Elder and Young, using the official task score once per task. Ours W. Acc. was not included in the available main-run summary and is left unreported.

Method	Macro-F1 $\uparrow$
Gemini 2.5 Flash	0.3236
MedAgent-Pro	0.7080
MERID (Ours)	0.7593

Appendix

Table 23: Kaggle raw-video emotion classification. Best and second-best scores are bold and underlined. MERID uses actor-disjoint nested cross-validation with review disabled.

Track	Window/task	Audio	Video	Max. length	Batch	Learning rate	Epochs
Elder	1-s binary	MFCC	OpenFace	26	1	$2\times 10^{-5}$	200
Elder	1-s ternary	OpenSMILE	ResNet	26	2	$4.5835899\times 10^{-6}$	400
Elder	1-s quinary	OpenSMILE	DenseNet	26	1	$1.6570181\times 10^{-5}$	400
Elder	5-s binary	OpenSMILE	ResNet	5	2	$1.8\times 10^{-5}$	200
Elder	5-s ternary	wav2vec	OpenFace	5	16	$1.4797182\times 10^{-4}$	400
Elder	5-s quinary	MFCC	ResNet	5	16	$9.8\times 10^{-5}$	200

Appendix

Table 24: Task-specific settings for the official MPDD 2025 reproduction.

Stage	Operation and retained consequence
Review	Roles analyze current feedback, prior outcomes, and active memory. The Chair issues a revision request or ends review.
Construct	The proposer turns a supported request into recipe or predictor additions. Source exclusions remove dependent candidates before reselection.
Evaluate	CPE evaluates untried pairs under the fixed protocol and remaining candidate budget. Each trial records predictions, its score, or an execution error.
Verify	The configured selection rule determines incumbent replacement. The round record retains the request and measured outcome.
Learn	Distillation proposes conditional lessons. Reconciliation adds, revises, or retires memory entries using the recorded evidence.
Re-enter	The next round receives the expanded space, retained pipeline, revision history, and updated memory. The loop ends on Chair acceptance or a budget limit.

Appendix

Table 25: Execution order of a memory-enabled improvement round. Learning also records nonpromoting and failed revisions.

Field	Meaning
id	Stable identity used to revise or retire an existing entry.
condition	Data and pipeline conditions under which the lesson applies.
lesson	An interpretation of the recorded experiment, marked as model synthesis.
next_action	A proposed experiment for extending or checking the interpretation.
evidence_ids	References to measured candidate records, round outcomes, or diagnostics.
status , versions	Active or retired status, with the previous contents retained after each update.

Appendix

Table 26: Research-memory fields. Numeric evidence remains attached to the underlying execution records.

Explore similar work

The MADRS Pipeline: Supporting Depression Assessment in Clinical Trials
Jul 30, 2026Mila Fodor, Katalin Ócsai, Francesco Periti +2Depression Mental Health
Clario, part of Thermo Fisher Scientific
Beyond Augmentation: Score-Guided Pathological Prior for EEG-based Depression Detection
May 29, 2026Xiaojing Chen, Jingqi Cheng, Xu Zhao +2Depression Electroencephalography
School of Computer Science and Technology, Hefei University of Technology Hefei, China · School of Computer Science and Information Engineering, Hefei University of Technology Hefei, China
EviDep: Uncertainty-Aware Multimodal Depression Estimation via Disentangled Evidential Learning
Apr 17, 2026Fangyuan Liu, Sirui Zhao, Yangsong Zhang +5Depression Evidential Uncertainty
School of Computer Science and Technology, University of Science and Technology of China, Hefei, China · aSchool of Computer Science and Technology, University of Science and Technology of China, Hefei, China · bSchool of Computer Science and Technology, Laboratory for Brain Science and Artificial Intelligence, Southwest University of Science and Technology, Mianyang, China +1

Jul 30, 2026, cs.CLJ/K move · Enter open · S save

The MADRS Pipeline: Supporting Depression Assessment in Clinical Trials

Mila Fodor, Katalin Ócsai, Francesco Periti, Rien Sonck+1

Clario, part of Thermo Fisher Scientific

Depression Mental Health Clinical Trials Cognitive Diagnosis Large Language Model Pipelines Assessment

Open paper AskPDF arXiv

Depression is a major mental disorder for which diagnosis relies primarily on clinical assessments. Automated methods to support its detection via the psychiatric MADRS scale are getting more and more attention. While existing solutions primarily focus on detecting the disorder from different text sources (e.g., online text, social media), there is still limited support for clinical trials, where clinical assessments are conducted through structured interviews based on standard guidelines such as SIGMA. In this work, we develop a LLM pipeline specifically designed to support clinicians in supporting the assessment of depression in patients enrolled in clinical trials. Our pipeline converts audio interviews into transcripts, maps them into the ten MADRS symptom items, estimates their severity, and identify problematic clinical ratings associated with them. Evaluation on real clinical interviews shows a strong overall correlation of 0.867 with expert ratings, providing interpretable support for future assessments in clinical trials.

Abstract

Explore similar work

The MADRS Pipeline: Supporting Depression Assessment in Clinical Trials

Beyond Augmentation: Score-Guided Pathological Prior for EEG-based Depression Detection

EviDep: Uncertainty-Aware Multimodal Depression Estimation via Disentangled Evidential Learning