acceptodds
Withdrawn from ICLR 2027

Agentic Adaptive AI for Clinical Care: Autonomous Construction of Patient-Specific Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We present an adaptive framework for personalized medicine designed to discover and deploy predictive models tailored to specific patients and their unique physiological dynamics. While static population models fail to capture individual physiological heterogeneity, standard patient-specific fine-tuning is not uniformly beneficial; sparse calibration budgets, high-frequency sensor noise, and severe domain shifts frequently degrade model accuracy below the original baseline, yet conventional systems lack mechanisms to assess whether adaptation should occur or when an unreliable model must be rejected. To address this, our architecture pairs an autonomous AutoResearch agent that iteratively searches across adaptation spaces (architectures, adapter ranks, and loss schedules) with a deterministic verification gate and a domain-specialized diagnostic agent. When a candidate breaches stability or physical invariants, the system triggers a diagnostic recovery loop, providing clinical oversight to pinpoint failure modes and safely arbitrate between matched cohort retrieval, training a scratch patient-only baseline, or refusing inference via data recollection. We evaluate our framework across two distinct clinical modalities: continuous ICU arterial blood pressure tracking from ECG and PPG waveforms (144 unseen PulseDB subjects), and simulation-to-real bioimpedance monitoring for dynamic fluid-shift tracking. Most notably, across both tasks, the framework cuts prediction error by approximately 50% relative to baseline model.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Personalized Medicine, Patient-Specific Modeling, Agentic AI, Adaptive Machine Learning, Failure-Aware Learning, Simulation-to-Real, AutoResearch Agents.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_GHhiOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces an AutoResearch framework for patient-specific physiological prediction. The pipeline contains an LLM search agent for exploring model adaptation strategies, a deterministic validation gate, and a failure-diagnosis LLM agent that provides feedbacks. The framework is evaluated on two applications, including blood pressure estimation from ECG/PPG signals, and bioimpedance-based fluid-shift estimation. Experiments show improvements over population-level models and standard adaptation baselines.

Critical strengths

None

Critical weaknesses
  1. Insufficient evidence to establish the necessity and effectiveness of the agentic components. The experiments do not demonstrate that the proposed multi-agent framework is necessary. For example: a) The failure-diagnosis agent provides limited benefits. In Table 1, the complete framework provides only marginal improvements over AutoResearch (SBP MAE: 5.77 → 5.72; DBP MAE: 3.32 → 3.28). In Figure 2C, AutoResearch already improves all 144 patients. b)The search scope of the AutoResearch agent is limited. The agent appears to operate primarily within a predefined search space for hyperparamters. It is unclear what advantages the LLM AutoResearch approach offers over conventional hyperparameter optimization or model selection methods. c) The framework is evaluated on only two small-scale datasets and it is unclear about its generalizability. In particular, the bioimpedance experiment uses real-world data from only two participants, which is insufficient to support claims about reliable patient-specific adaptation.

  2. Poor presentation with limited scientific insights. The paper reads more like a case study report than a research paper and provides little scientific insight.

Questions for authors

None

Reviewer_chNmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents an agentic framework for adapting physiological prediction models to individual patients. It combines LLM-guided adaptation search with a validation gate and a diagnostic router that selects recovery strategies when adaptation fails. Evaluations on ECG/PPG-based blood-pressure estimation for 144 held-out PulseDB subjects and simulation-to-real bioimpedance volume-shift estimation for two volunteers show reduced prediction errors relative to the evaluated population and fixed-adaptation baselines.

Critical strengths

The paper explores a practical problem. Personalization does not always help, and a simpler patient-only model may work better when adaptation fails.

The PulseDB evaluation separates model search, routing, and final testing on 144 held-out subjects. The improvements over population models and fixed personalization are encouraging.

Critical weaknesses

The bioimpedance study involves only two volunteers, and several important details are unclear, including how volume changes were measured, how repeated measurements were split, and how many labels were used for validation and routing. This makes the reported results difficult to assess.

The added benefit of the diagnostic router, particularly its safety benefit, is unclear. On PulseDB, it only slightly improves over AutoResearch, which already improves predictions for all 144 patients at the largest calibration budget. The paper also does not establish how reliably the gate detects harmful adaptations or whether abstention decisions are appropriate.

Questions for authors

What domain expertise is required to build the pipeline, particularly the diagnostic agent and its recovery rules? Please clarify how this knowledge is provided to the agent and how its diagnostic recommendations are validated. This would help distinguish what the agent discovers from what is specified by the researchers.

For the bioimpedance experiment, how were the reference volume changes measured, and how were participants, sessions, and repeated measurements divided between calibration, validation, and testing?

Reviewer_jxXxOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a failure-aware agentic framework for patient-specific adaptation of physiological prediction models.The framework uses an LLM-based AutoResearch agent to search for patient-specific model configurations and is evaluated on two tasks: continuous blood pressure estimation and simulation-to-real bioimpedance-based fluid shift estimation.

Critical strengths

The paper addresses the risk of negative transfer, the evaluation includes two different physiological prediction tasks.

Critical weaknesses
  1. The paper focuses on transfer learning and model adaptation (as positioned in primary area), but the methodological contribution is limited. The proposed framework mainly combines existing methods, such as fine-tuning, hyperparameter optimization, and rule-based routing.
  2. The experiments do not clearly show whether LLM agents provide additional benefits over conventional HPO methods and simpler routing strategies.
  3. Case 2 involves only two healthy participants, making it difficult to assess generalizability.
  4. The multi-stage framework introduces additional computational costs through repeated model optimization and agent-based routing. The paper does not evaluate these costs or discuss whether the framework is practical for clinical deployment.
Questions for authors

Case 1: 1.AutoResearch strategy:
a. Does AutoResearch iteratively update the configurations based on previous validation results, or are all four candidates generated in a single LLM call? b. Table 1 shows AutoResearch outperforms Random HPO, how about the performance compared with Bayesian optimization and other improved HPO?

2.Failure-aware routing. Appendix A.2 clarifies the deterministic failure gate and the information provided to the diagnostic agent. a.How the agent distinguishes different failure modes and selects an appropriate recovery strategy? b. What is the optimal routing accuracy? And whether LLM-based diagnosis provides additional benefits? c. What is the negative transfer?

3.Clinical validity. Please clarify the feasibility of obtaining reference BP labels for patient-specific calibration and assess the reliability of routing decisions based on only 20 validation windows.

Case 2:

  1. Limited real-world evaluation. a. The real-world evaluation includes only two healthy participants. b. How are calibration, validation, and test samples split across sessions, electrode placements, and volume-shift levels?
  2. Ground-truth volume shifts. a. How are the 11 nominal volume-shift targets obtained? b. How is the ground truth validated?
  3. Sim-to-real adaptation. a. The authors generated over one million synthetic spectra but used only 100,000 for pretraining. What is the rationale for this selection? b. How does the patient-only rescue model perform compared with synthetic-pretrained adaptation?
  4. Failure-aware routing. a. Similar concern as Case 1. b. How were the gate thresholds determined, and how reliable are they with sparse real-world calibration data?

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

Related papers

Discussion (0)

Sign in to comment.