acceptodds
Withdrawn from ICLR 2027

InertiaMPDD: Multimodal Personalized Depression Detection via Individual Emotional-Inertia Modeling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing omni-modal depression detection largely infers depression representations shared across individuals, rather than latent variables of individual difference inferred from each person's behavioral signals. Yet depressive expression superimposes transient states on stable individual traits: unless traits are modeled and conditioned on, they confound the assessment of states, while conditioning on annotated individual attributes does not scale because such annotations are scarce. We propose InertiaMPDD, which reframes detection as estimating each person's deviation of current expression from a counterfactual healthy baseline, turning individual differences from externally supplied labels into latent variables inferred from behavioral signals. The baseline is an individual-conditioned latent dynamics predictor trained exclusively on healthy dynamics, so departures from healthy expression systematically inflate its prediction error. Rather than regenerating a "healthy" counterpart, we realize the counterfactual by single-step rollout: under the person's own trait conditioning, the model advances the observed state one step according to healthy dynamics, and the deviation constitutes the individualized depressive signal. Because prediction and observation share the same trajectory, context and recording conditions, individual traits appear on both sides and cancel. Three complementary, annotation-free quantities follow: the residual (position), the AR(1) restoring force (speed of return to baseline), and the spectral radius of the predictor's Jacobian at the individual operating point (local stability of healthy dynamics). On the in-the-wild vlog corpora D-Vlog and LMVD, InertiaMPDD attains F1 of 98.41 and 84.98, above the best published results on both, and the joint system surpasses either pathway alone. The residual is significant with a consistent direction on both datasets.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Depression Detection, Multimodality, Personalization, LLM

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_crKjOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This article proposed a new method called InertiaMPDD, for depression detection. InertiaMPDD frames detection as estimating each person’s deviation of current expression from a counterfactual healthy baseline, turning individual differences from externally supplied labels into latent variables inferred from behavioral signals. The baseline is an individual-conditioned latent dynamics predictor trained exclusively on healthy dynamics. D-VLOG and LMVD were used for testing.

Critical strengths
  1. The statistical analyses are comprehensive.
  2. The idea of following each individual's healthy baseline and tracing individual change makes sense.
Critical weaknesses
  1. A fundamental flaw of the proposed method is assuming that a deviation from an individual's norm must be depression. This is rarely true in real life or even within the D-Vlog and LMVD datasets, since depression being assessed in the datasets does not mean that the individuals did not have other mental/behavioral concerns. These good results could be due to the fact that only depression was included in the datasets as the possible deviation. That means, as long as the dataset contains other deviations, the method could fail. It was unclear if the uniqueness of depression was effectively modeled.
  2. How "healthy" and "depressive" behaviors are labeled in the used datasets is very important to understand if the proposed method would work at least on these datasets. However, this was not clearly explained; thus, how the data were parsed and marked was unknown.
  3. The formulation does not increase the understanding of the methods since how each modality was handled and balanced, as well as which part is unique to depression was still unclear.
Questions for authors
  1. How was the sensitivity to depression evaluated empirically?
Reviewer_xGstOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes InertiaMPDD, a multimodal depression detection framework combining a Qwen2.5-Omni screening pathway with a latent dynamics predictor trained on healthy samples. The predictor is conditioned on recording-level statistics and produces three features based on prediction residuals, AR(1) decay, and the spectral radius of its Jacobian. These features adjust the screening margin through a supervised bias head. The method reports F1 scores of 98.41% on D-Vlog and 84.98% on LMVD.

Critical strengths
  1. Combining the screening and dynamics pathways improves the reported F1 over either pathway alone, particularly on LMVD, where the combined system achieves 84.98% compared with 74.57% for NSP and 75.88% for IIP.
  2. The paper examines prediction residuals, autoregressive decay, and Jacobian spectra separately.
Critical weaknesses
  1. Recent methods evaluated on the same D-Vlog and LMVD corpora, including MMFformer (SMC'25), CAF-Mamba (ICASSP'26), DynMultiDep (Journal of Imaging 26), ASYM (Frontiers in Psychiatry 26), AttMamba (Neurocomputing 26), and BHDD (IEEE TAFFC 26), have already investigated temporal dependencies, adaptive multimodal fusion, and behavioral representations for depression detection. PAVNet (BIBM 25) has also explored data-driven personality modeling through adaptive clustering and longitudinal subject information, although on a different corpus. While these methods do not reproduce the proposed healthy-only conditional predictor, they substantially narrow the claimed novelty. Learning normal dynamics and detecting deviations through prediction errors are also established principles in anomaly detection. The potentially distinctive contribution is therefore the combination of a healthy-trained, trait-conditioned predictor with supervised screening. However, the inferred trait representation comes from a single recording, with no evidence that it captures stable individual characteristics rather than transient depressive states or recording conditions. Crucial comparisons against unconditioned predictors, shuffled trait representations, generic prediction errors, and conventional margin calibration are missing. Consequently, the experiments do not establish that the improvements originate from personalized healthy dynamics rather than established temporal modeling and supervised feature fusion.
  2. Proposition 1 is a standard squared-error decomposition and does not establish that prediction residuals specifically reflect depression rather than noise or recording differences. Propositions 2 and 3 require common state coordinates and appropriate dynamical assumptions. In the actual architecture, however, the predictor maps online representations to a separately encoded target space. The spectral radius of this mapping need not correspond to the relaxation rate of the underlying emotional process and can change with the target coordinate system. Moreover, the recording-level trait representation is computed from the full sequence and may contain future information relative to individual prediction steps. Thus, the rollout is not strictly causal. The estimation of a scalar AR(1) coefficient from 256-dimensional trajectories is also insufficiently specified.
  3. The D-Vlog results in Table 1 are consistent with a 175-sample evaluation subset, whereas the original corpus publication reports 212 test samples. The difference is not explained. Furthermore, the original D-Vlog baseline reports weighted metrics, while the proposed method reports positive-class metrics, making direct comparisons potentially inconsistent. The provenance of the raw audio, video frames, and transcripts is also unclear because the original corpora do not provide all these modalities in the form required by Qwen2.5-Omni. The near-ceiling NSP performance on D-Vlog further requires text-only and diagnostic-language masking controls to establish whether the model learns behavioral indicators or primarily exploits explicit self-reports. The combined system improves D-Vlog F1 by only 1.03 percentage points, corresponding to a net reduction of two false positives, without a paired significance test. LMVD results are averaged over five seeds without reporting standard deviations or confidence intervals.
  4. The supervised InertiaMPDD results are compared with DepressionAgent, which uses a training-free evaluation setting, without controlling for the different supervision and computational requirements.
Questions for authors
  1. What are the results when removing or shuffling trait conditioning, replacing healthy-only training with conventional sequence prediction, and replacing IIP with generic latent statistics or supervised margin calibration?
  2. Can cross-recording stability of the inferred traits be demonstrated?
  3. How is the Jacobian spectral radius justified as a dynamical stability measure when online and target representations use different encoders?
  4. How is the scalar AR(1) coefficient estimated from multidimensional trajectories, and how is future information excluded from recording-level conditioning?
  5. Can the authors provide paired significance tests for NSP versus InertiaMPDD, per-seed LMVD results with confidence intervals, and text-only and diagnostic-language masking experiments?
  6. How much additional computation does the complete IIP pathway require?
Reviewer_WRZ2OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

Prior works on depression detection usually make training and testing on a dataset-level, so every subject is treated as by the labels. This work propose that there are individual differences, such as demographic factors, that are ignored, thus make the prediction less useful and reliable,

So they try to build representations from a complete healthy group, and measure how each subject deviated from the healthy representation. that is , from each subject's own healthy representation. So the prediction is more personalized, other than cohort wise.

Critical strengths

The focus on individualized prediction makes sense, as depressive symptoms vary significantly across different populations. So this research direction is promising.

Critical weaknesses
  1. The presentation of the manuscript needs a lot of improvement. When Section 2 introduces the Methodology, the essential mathematical notations are relegated to the Appendix. This separation makes the technical details difficult to follow and disrupts the flow of the text.

  2. The description of the datasets is insufficient. Following a very brief introduction of the data, the authors immediately transition to the experimental setup and results. Crucial details—such as the data modalities, distributions, and sample sizes—need to be elaborated much more clearly to provide adequate context.

  3. The formulation of the "healthy cohort" remains unclear, despite being a primary point of interest. It is not apparent whether each subject is matched with a corresponding healthy control sample. The authors should use accessible language to explain precisely how this cohort is constructed and what ensures its reliability. Introducing complex terms like "healthy dynamics" and "world model" without sufficient explanation further obfuscates the core methodology.

Questions for authors

To the best of my knowledge, the raw video files of the D-Vlog dataset cannot be publicly shared due to privacy regulations. Could the authors clarify exactly how the data was accessed and processed for training and testing?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.