Pith. sign in

REVIEW 4 major objections 8 minor 70 references

Smarter Together: Combining Large Language Models and Small Models for Physiological Signals Visual Inspection

T0 review · 4 major / 8 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ConMIL turns a small conformalized MIL model into a plug-in that lifts frozen LLMs' ECG/EEG reading from 13-48% to ~95% on confident samples.

desk verdict Solid empirical hybrid-SSM/LLM pipeline; the per-class conformal guarantee is contradicted by the paper's own Table A2, so treat the reliability claim as unproven. read the letter →

arxiv 2501.16215 v2 pith:W73CHX56 submitted 2025-01-27 cs.AI cs.LGeess.SP

Huayu Li , Zhengxiao He , Xiwen Chen , Ci Zhang , Stuart F. Quan , William D.S. Killgore , Shu-Fen Wung , Chen X. Chen
show 3 more authors
Geng Yuan Jin Lu Ao Li
This is my paper · ORCID
classification cs.AIcs.LGeess.SP
keywords physiologicalsignalclassificationmultipleinstancelearningconformalpredictionlargelanguagemodelsvisualinspectionECGarrhythmiadetectionsleepstageinterpretability
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that the weak spot of large language models on medical waveforms—domain-specific precision—can be fixed without fine-tuning the LLM, by feeding it calibrated hints from a small specialized model. Its ConMIL framework trains a multiple-instance model to say which time segments support each candidate class, then conformalizes those scores so the output is a set of diagnoses with a per-class coverage guarantee. On arrhythmia and sleep-stage data, frozen Qwen2-VL-7B and MiMo-VL-7B-RL supported by ConMIL reach 94.92% and 96.82% precision on confident (singleton) samples, compared with 46.13% and 13.16% when the LLM reads the plot alone. If this holds, it gives a recipe for trustworthy visual inspection of physiological signals that keeps the small model's specialization and the LLM's reasoning, with ambiguous cases explicitly flagged for human review.

What carries the argument

The load-bearing object is QTrans-Pooling, a pooling layer that gives each of the K classes its own learnable token and computes class-specific attention over time instances through sigmoid cross-attention with bias -log(T). This produces per-class heatmaps that say which time segments support each possible diagnosis. On top of it, conformal risk control calibrates per-class thresholds on a held-out set so that the set-valued prediction contains the true class with probability at least 1-alpha for each class; the decision rules then map singleton, multi-label, and trivial sets to direct acceptance, LLM adjudication, or clinician referral.

What would settle it

On a fresh exchangeable split of PTB-XL or SleepEDF with alpha=0.05, compute per-class false-negative rates from ConMIL's set-valued predictions; if any class exceeds 5%, as the paper's own supplementary table shows for some classes, the per-class coverage guarantee is not demonstrated.

Watch

Extended reading notes

Core claim

ConMIL is a decision-support framework that combines three things: a per-class interpretable MIL pooling layer called QTrans-Pooling, conformal risk control for set-valued predictions with a per-class false-negative guarantee, and a prompting scheme that hands both the prediction set and per-class heatmaps to a frozen multimodal LLM. The paper's central discovery is that this combination converts an LLM's visual inspection of ECG and EEG from a mostly unreliable guess into a stratified, evidence-based decision: singleton sets are read with high confidence, multi-label sets are adjudicated using per-class interpretations, and trivial sets that include all classes are sent to a clinician. Empirically, this support lifts Qwen2-VL-7B from 46.13% accuracy on PTB-XL and 13.16% on SleepEDF to 94.92% and 96.82% on confident samples, with 70.61%/78.10% and 78.02%/71.98% on uncertain samples, and the same pattern holds for MiMo-VL-7B-RL. ConMIL is described as the first combination of conformal prediction with MIL, and the per-class coverage guarantee is the formal basis for the reliability claims.

Load-bearing premise

The reliability guarantee depends on the calibration and test data being exchangeable and on per-class thresholds actually capping each class's miss rate at the chosen level.

Editorial extensions

If this is right

  • Both Qwen2-VL-7B and MiMo-VL-7B-RL reach 94.92% precision on confident PTB-XL samples and 96.82% on confident SleepEDF samples, against 46.13%/48.16% and 13.16%/21.08% without support.
  • On uncertain multi-label samples the same LLMs score 70.61%/78.10% on PTB-XL and 78.02%/71.98% on SleepEDF, still well above their standalone baselines.
  • Instead of forcing a guess, ConMIL can defer: at alpha=0.05, 171 PTB-XL samples are rejected for clinician review, while SleepEDF produces no trivial sets.
  • Because the LLM is used frozen, the framework works with proprietary models that cannot be fine-tuned, avoiding the main barrier to domain adaptation.
  • Per-class heatmaps are what let the LLM separate visually similar classes such as N2 versus REM or ST/T change versus hypertrophy; prediction sets alone, as in the Monty Hall comparison, perform far worse.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If ConMIL generalizes, the same plug-in design could be applied to other waveform-reading tasks, such as fetal heart-rate monitoring or intraoperative EEG, whenever a small model can output per-class saliency; the test would be whether the heatmaps remain discriminative for classes that are not mutually exclusive.
  • A natural deployment is a triage protocol where singleton outputs auto-approve, multi-label outputs go to LLM-assisted review, and trivial outputs go straight to a clinician; the paper's numbers imply most SleepEDF cases are singleton at alpha=0.05, so workload savings could be large.
  • If per-class heatmaps are the active ingredient, a controlled ablation that keeps the set-valued output but removes the heatmaps would quantify how much of the gain comes from interpretability rather than from mere narrowing of the label set.
  • Under distribution shift, the non-exchangeable bound in the paper means coverage degrades with the total variation distance between calibration and deployment data; in practice this points to periodic local recalibration before using ConMIL in a new hospital or device population.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 8 minor

Summary. The paper introduces ConMIL, a decision-support framework that combines multiple instance learning with per-class interpretability (QTrans-Pooling), conformal risk control for set-valued predictions, and prompting of multimodal LLMs for visual inspection of physiological signals. The authors evaluate ConMIL on SleepEDF and PTB-XL using Qwen2-VL-7B, MiMo-VL-7B-RL, and ChatGPT-4.0, reporting large accuracy gains on confident and uncertain samples relative to standalone LLM use, and they provide qualitative case studies and an 'interview' protocol. The main advertised contribution is twofold: per-class interpretable MIL pooling, and a conformalized MIL pipeline that gives per-class false-negative-rate guarantees under exchangeability.

Significance. If the claims hold, the framework is a valuable practical contribution: it combines the interpretability and uncertainty quantification of a small specialized model with the reasoning of a frozen multimodal LLM, without fine-tuning the LLM, and it releases code and uses public datasets. The empirical gains reported are large and the qualitative material is detailed. However, the central statistical-reliability claim is not supported by the paper's own numbers: Table A2 reports per-class FNRs above the promised α at α=0.05, and the proof of per-class control is not supplied. The LLM evaluation also has reproducibility and confound issues. The overall direction is promising and the empirical pipeline is clearly described, but the reliability guarantee needs repair or substantial qualification.

major comments (4)
  1. [Section 3.3, Eq. (15)-(16), Theorem 3, and Table A2] The headline guarantee is per-class coverage: P(y_test in S_α(X_test) | y_test = k) ≥ 1-α for every k. Table A2 directly contradicts this at α=0.05: PTB-XL class 2 has per-class FNR 6.31±0.65% and SleepEDF class 5 has FNR 10.68±0.71%, both above 5% by amounts far exceeding the finite-sample slack B/(n+1). The paper reports marginal coverage (94.65% and 95.59%) that meets α, but marginal coverage is not the advertised per-class guarantee. Since the abstract and Section 3.3 sell the framework precisely as providing per-class statistical guarantees, this internal inconsistency undermines the central reliability claim and must be fixed before publication.
  2. [Section 3.3, Eq. (14) and Definition 1] The Conformal Risk Control theorem quoted from [38] applies to a single scalar threshold λ with a bounded, right-continuous, non-increasing risk function. Here the prediction set is built from K class-specific thresholds λ_k and the risk in Eq. (12) is class-conditioned. The manuscript does not prove that CRC applied to K thresholds, or to class-conditioned calibration subsets, yields per-class FNR control. The empirical violation in Table A2 suggests this is not just a missing proof but a real gap. The authors should either provide a class-wise calibration argument with explicit finite-sample bounds, or replace the claim with a marginal-coverage guarantee and state that per-class control is not guaranteed.
  3. [Section 5.3, Figures 3-4 and decision rules in Section 4.2] The LLM evaluation is confounded by the task asymmetry. In the standalone condition the LLM performs an open 5-way classification; in the ConMIL condition it receives a singleton or a very small set (e.g., two classes) plus per-class heatmaps and is instructed to choose 'without additional analysis.' The large accuracy jump on confident samples may therefore largely reflect ConMIL's own accuracy and the reduced label space, rather than the LLM's visual inspection ability. To support the claim that ConMIL 'enhances visual inspection capabilities,' the authors should report per-sample comparisons, including cases where the ConMIL set contains the true label but the singleton/uncertain split is controlled, and should report the ConMIL-only accuracy on the same confident/uncertain subsets. Without this, the contribution to LLM reasoning, as opposed to providing a nearly solved subset, is not established.
  4. [Section 5.2 and Section 5.3] All LLM results are single-run, and ChatGPT-4.0 is accessed through the OpenAI GUI, which is not reproducible and may change over time. For a central empirical claim of 94.92% vs. 46.13%, the paper should provide confidence intervals, multiple seeds or repeated API calls, temperature settings, and the exact model versions. The qualitative 30-case study in Table 1 is useful but too small to carry the quantitative weight of the comparison.
minor comments (8)
  1. [Abstract] The phrase 'distinctively synergizes' is ungrammatical and should be revised.
  2. [Section 3.3, Eq. (12)] The sum in the FNR definition runs from i=0 to N_k; it should run from i=1 to N_k.
  3. [Section 3.3, Eq. (13) and Eq. (14)] The notation is confusing: Eq. (13) defines a binary prediction ŷ_i using a threshold λ, while Eq. (14) reuses λ as a vector of K thresholds and writes p(X_i) where the per-class probability p_k(X_i) is intended. Please clarify the notation and distinguish scalar and vector thresholds.
  4. [Section 3.2, Theorem 2] Theorem 2 is essentially the elementary fact that conditioning cannot increase entropy. Calling this a reduction in class-wise variability that 'simplifies decision boundaries' overstates what the inequality establishes; please either provide a more substantive statement or present it as a heuristic remark.
  5. [Section 2.4] There is a typo: 'MModels like Med-PaLM' should be 'Models like Med-PaLM.'
  6. [Section 2.3] The phrase 'per-class interpretabilityg' contains a stray 'g' and needs correction.
  7. [Section 4.2 and Figure 2] The decision rule for singleton predictions says the LLM 'interprets the result directly,' but the paper does not specify whether the LLM is allowed to disagree with a singleton ConMIL prediction. This matters for interpreting the confident-sample accuracy numbers.
  8. [Section 3.1, Eq. (2)] The majority-voting rule y_i = argmax_k Σ_t y_{i,k} is formally undefined for ties and is not used in the later conformal set construction; please reconcile this with the set-valued prediction definition.

Circularity Check

1 steps flagged · score 4.0 of 10

One headline metric (confident-sample LLM accuracy) is forced by the singleton decision rule; the conformalized-MIL core and uncertain-sample results are independent.

  1. fitted input called prediction [Section 4.2 (Singleton Prediction) and Figure 2(B); results in Section 5.3 and Abstract]
    "Singleton Prediction (High Confidence) When ConMIL generates a singleton prediction (a single class in the prediction set), LLMs interpret the result directly using the associated interpretability prompts. The (1−α)% confidence guarantee ensures high reliability, enabling the generation of precise clinical recommendations with minimal uncertainty. ... ConMIL-supported Qwen2-VL-7B and MiMo-VL-7B-RL both achieves 94.92% and 96.82% precision on confident samples"

    In the singleton branch, the LLM is instructed to adopt the single class contained in the prediction set (Eq. 14 with |S|=1); the prompt offers a one-element choice. The reported 'LLM accuracy on confident samples' is therefore, by construction, ConMIL's own accuracy on the high-confidence subset, not an independent LLM visual-inspection prediction. Comparing this number to the LLM-alone accuracy of 46.13%/13.16% over the full test set contrasts a forced echo of ConMIL's singleton with an unassisted LLM decision. The uncertain-sample accuracies (70-78%) do involve genuine LLM selection among multiple options, so the central claim retains independent support; this circularity is partial and confined to the headline confident-sample statistic.

full rationale

The conformalized-MIL derivation is self-contained: QTrans-Pooling is an architectural extension of TimeMIL [5] (co-authored, but the pooling mechanism and per-class tokens are new), and the CRC guarantee in Theorem 3 is quoted from the external Angelopoulos et al. [38] work, not from a self-citation. No fitted parameter is relabeled as a prediction in the conformal step; thresholds are calibrated on the validation set and evaluated on held-out test data. The per-class FNR violation visible in Table A2 (PTB-XL class 2: 6.31% at alpha=0.05; SleepEDF class 5: 10.68%) is a correctness/validity gap: Eq. 16 and Theorem 3 are stated for a scalar lambda and a monotone bounded risk, and the paper does not prove that K class-specific thresholds with class-conditioned FNR risks inherit the guarantee. This is a failure of the stated assumptions, not a circular reduction, so it does not raise the circularity score. The only circular step found is the singleton-decision metric, where 'LLM accuracy on confident samples' equals ConMIL's singleton precision by the decision rule in Section 4.2/Figure 2B; the abstract's comparison with the LLM-alone full-set accuracy is therefore not a like-for-like measure of LLM improvement. Apart from this forced headline statistic, the uncertain-sample evaluations and the external baselines provide independent content.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The main load-bearing ingredients are statistical assumptions (exchangeability, monotone risk) and imported theorems. No new physical entities are introduced. The conformal thresholds are calibrated parameters, not first-principles constants.

free parameters (2)
  • Per-class conformal thresholds lambda_k = Selected via CRC on the validation set; exact values not reported in the paper
    Determines singleton, uncertain, and reject splits and therefore the reported accuracies; fitted to calibration data (Section 3.3, Eq. 16).
  • Sigmoid attention bias b = -log(T) = b = -log(T), where T is the number of instances
    Hand-chosen bias in Eq. (10) to control attention saturation; a design parameter, not fitted to data.
assumptions (5)
  • domain assumption Calibration and test data are exchangeable
    Needed for Theorem 3 and Eq. 16 coverage guarantee; Section 3.3 Definition 2. Table A2 suggests the resulting per-class coverage can fall below the nominal level.
  • domain assumption CRC risk function R_lambda is right-continuous, non-increasing in lambda, and bounded
    Required to apply Conformal Risk Control (Theorem 3); Section 3.3 Eq. 15-16.
  • standard math MIL universal approximation theorem for symmetric, continuous score functions
    Used to justify the MIL pooling architecture in Section 3.2 Theorem 1; imported from [33,34].
  • standard math Conditioning on additional random variables cannot increase entropy
    Used in Theorem 2, Eq. 11; the interpretation that this guarantees simpler decision boundaries is not proven.
  • standard math TV-distance bound for non-exchangeable conformal prediction
    Imported from [39] to extend guarantees under distribution shift; only bounds the added coverage gap, does not eliminate it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Smarter Together: Combining Large Language Models and Small Models for Physiological Signals Visual Inspection." pith.science (2026). https://pith.science/paper/W73CHX56

@misc{pith2026250116215,
  author       = {Pith},
  title        = {Pith review of: Smarter Together: Combining Large Language Models and Small Models for Physiological Signals Visual Inspection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W73CHX56}},
  note         = {Machine review of arXiv:2501.16215}
}
read the original abstract

Large language models (LLMs) have shown promising capabilities in visually interpreting medical time-series data. However, their general-purpose design can limit domain-specific precision, and the proprietary nature of many models poses challenges for fine-tuning on specialized clinical datasets. Conversely, small specialized models (SSMs) offer strong performance on focused tasks but lack the broader reasoning needed for complex medical decision-making. To address these complementary limitations, we introduce \ConMIL{} (Conformalized Multiple Instance Learning), a novel decision-support framework distinctively synergizes three key components: (1) a new Multiple Instance Learning (MIL) mechanism, QTrans-Pooling, designed for per-class interpretability in identifying clinically relevant physiological signal segments; (2) conformal prediction, integrated with MIL to generate calibrated, set-valued outputs with statistical reliability guarantees; and (3) a structured approach for these interpretable and uncertainty-quantified SSM outputs to enhance the visual inspection capabilities of LLMs. Our experiments on arrhythmia detection and sleep stage classification demonstrate that \ConMIL{} can enhance the accuracy of LLMs such as ChatGPT4.0, Qwen2-VL-7B, and MiMo-VL-7B-RL. For example, \ConMIL{}-supported Qwen2-VL-7B and MiMo-VL-7B-RL both achieves 94.92% and 96.82% precision on confident samples and (70.61% and 78.02%)/(78.10% and 71.98%) on uncertain samples for the two tasks, compared to 46.13% and 13.16% using the LLM alone. These results suggest that integrating task-specific models with LLMs may offer a promising pathway toward more interpretable and trustworthy AI-driven clinical decision support.

Figures

Figures reproduced from arXiv: 2501.16215 by the authors.

Figure 1
Figure 1. Illustration of the MIL model used in ConMIL. Our model is adapted from the one proposed by [5] which uses InceptionTime [3] as the feature extractor and encodes temporal correlations with Wavelet Positional Encoding (WPE) [5]. Instead of Trans-Pooling, we have introduced the QTrans-Pooling for enhanced per-class interpretability. 3.2 QTrans-Pooling with per-class interpretability We first recap the approximation of… view at source ↗
Figure 2
Figure 2. Overview of the ConMIL workflow for enhancing LLM-driven visual inspection of physiological signals. (A) Training and Calibration Process: ConMIL undergoes supervised training with QTrans-Pooling for per-class interpretability and calibration with conformal prediction to generate reliable thresholds (λ1, . . . , λK). This ensures set-valued predictions with guaranteed confidence levels. (B) Decision Support Framewor… view at source ↗
Figure 3
Figure 3. ConMIL on Qwen2-VL-7B’s predictions for PTB-XL (A) and SleepEDF (B) at α = 0.05. Qwen2- VL-7B, supported by ConMIL, achieves 94.92% and 96.82% accuracy for confident samples and 70.61% and 78.02% for uncertain samples, with 171 and 0 samples rejected, respectively. This performance is notably higher than the LLM operating alone, with MILLET support, or with the Monty Hall method, highlighting ConMIL’s effectiveness … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: ConMIL on MiMo-VL-7B-RL’s predictions for PTB-XL (A) and SleepEDF (B) at α = 0.05. The framework shows consistent benefits with a different LLM. Supported by ConMIL, MiMo-VL-7B-RL attains an accuracy of 94.92% on confident samples for PTB-XL and 96.82% for SleepEDF. Ac…
Figure 5
Figure 5. Figure 5: Case studies demonstrating how ConMIL enhances ChatGPT4.0’s capabilities in sleep stage classification and arrhythmia classification. (A) Sleep Stage Classification: ChatGPT4.0 correctly identifies the N2 stage only when supported by ConMIL’s set-valued prediction and …
Figure 6
Figure 6. Figure 6: “Interview” of ChatGPT4.0 on EEG sleep stage classification (A) and ECG arrhythmia classification [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 59 canonical work pages

  1. [38]

    Conformal risk control.arXiv preprint arXiv:2208.02814, 2022

    Anastasios N Angelopoulos, Stephen Bates, Adam Fisch, Lihua Lei, and Tal Schuster. Conformal risk control.arXiv preprint arXiv:2208.02814, 2022. 3.3, 1, 3.3

    arXiv 2022

  2. [1]

    Benjamin Shickel, Patrick James Tighe, Azra Bihorac, and Parisa Rashidi. Deep ehr: a survey of recent advances in deep learning techniques for electronic health record (ehr) analysis.IEEE journal of biomedical and health informatics, 22(5):1589–1604, 2017. 1

    work page 2017

  3. [2]

    Recurrent neural networks for multivariate time series with missing values.Scientific reports, 8(1):6085, 2018

    Zhengping Che, Sanjay Purushotham, Kyunghyun Cho, David Sontag, and Yan Liu. Recurrent neural networks for multivariate time series with missing values.Scientific reports, 8(1):6085, 2018. 1

    work page 2018

  4. [3]

    Inceptiontime: Finding alexnet for time series classification.Data Mining and Knowledge Discovery, 34(6):1936–1962, 2020

    Hassan Ismail Fawaz, Benjamin Lucas, Germain Forestier, Charlotte Pelletier, Daniel F Schmidt, Jonathan Weber, Geoffrey I Webb, Lhassane Idoumghar, Pierre-Alain Muller, and Fran¸ cois Petitjean. Inceptiontime: Finding alexnet for time series classification.Data Mining and Knowledge Discovery, 34(6):1936–1962, 2020. 1, 2.1, 1, A, A1

    work page 1936

  5. [4]

    Medformer: A multi-granularity patching transformer for medical time-series classification.arXiv preprint arXiv:2405.19363, 2024

    Yihe Wang, Nan Huang, Taida Li, Yujun Yan, and Xiang Zhang. Medformer: A multi-granularity patching transformer for medical time-series classification.arXiv preprint arXiv:2405.19363, 2024. 1, 2.1, A, A1

    arXiv 2024

  6. [5]

    TimeMIL: Advancing multivariate time series classification via a time-aware multiple instance learning

    Xiwen Chen, Peijie Qiu, Wenhui Zhu, Huayu Li, Hao Wang, Aristeidis Sotiras, Yalin Wang, and Abolfazl Razi. TimeMIL: Advancing multivariate time series classification via a time-aware multiple instance learning. InForty-first International Conference on Machine Learning, 2024. 1, 2.2, 1, 3.2, 3.2, 3.2, 2, A, A1

    work page 2024

  7. [6]

    Inherently interpretable time series classification via multiple instance learning

    Joseph Early, Gavin Cheung, Kurt Cutajar, Hanting Xie, Jas Kandola, and Niall Twomey. Inherently interpretable time series classification via multiple instance learning. InThe Twelfth International Conference on Learning Representations, 2024. 1, 2.2, 3.2, A, A1

    work page 2024

  8. [7]

    Can generalist foundation models outcompete special-purpose tuning? case study in medicine.arXiv preprint arXiv:2311.16452, 2023

    Harsha Nori, Yin Tat Lee, Sheng Zhang, Dean Carignan, Richard Edgar, Nicolo Fusi, Nicholas King, Jonathan Larson, Yuanzhi Li, Weishung Liu, et al. Can generalist foundation models outcompete special-purpose tuning? case study in medicine.arXiv preprint arXiv:2311.16452, 2023. 1

    arXiv 2023

Show all 70 references
  1. [8]

    Towards expert-level medical question answering with large language models.arXiv preprint arXiv:2305.09617, 2023

    Karan Singhal, Tao Tu, Juraj Gottweis, Rory Sayres, Ellery Wulczyn, Le Hou, Kevin Clark, Stephen Pfohl, Heather Cole-Lewis, Darlene Neal, et al. Towards expert-level medical question answering with large language models.arXiv preprint arXiv:2305.09617, 2023. 1, 2.4

    2023 arXiv

  2. [9]

    Sleep stage estimation by introduction of sleep domain knowledge to ai: Towards personalized sleep counseling system with genai

    Iko Nakari and Keiki Takadama. Sleep stage estimation by introduction of sleep domain knowledge to ai: Towards personalized sleep counseling system with genai. InProceedings of the AAAI Symposium Series, volume 3, pages 368–373, 2024. 1, 2.4, 5.4

    2024

  3. [10]

    Serkan G¨ unay, Ahmet¨Ozt¨ urk, and Yavuz Yi˘ git. The accuracy of gemini, gpt-4, and gpt-4o in ecg analysis: A comparison with cardiologists and emergency medicine specialists.The American journal of emergency medicine, 84:68–73, 2024. 1, 2.4, 5.4

    2024

  4. [11]

    Comparison of emergency medicine specialist, cardiologist, and chat-gpt in electrocardiography assessment.The American Journal of Emergency Medicine, 80:51–60, 2024

    Serkan G¨ unay, Ahmet ¨Ozt¨ urk, Hakan¨Ozerol, Yavuz Yi˘ git, and Ali Kemal Erenler. Comparison of emergency medicine specialist, cardiologist, and chat-gpt in electrocardiography assessment.The American Journal of Emergency Medicine, 80:51–60, 2024. 1

    2024

  5. [12]

    Multimodal chatgpt-4v for electrocardiogram interpretation: Promise and limitations.Journal of Medical Internet Research, 26:e54607, 2024

    Lingxuan Zhu, Weiming Mou, Keren Wu, Yancheng Lai, Anqi Lin, Tao Yang, Jian Zhang, and Peng Luo. Multimodal chatgpt-4v for electrocardiogram interpretation: Promise and limitations.Journal of Medical Internet Research, 26:e54607, 2024. 1, 5.4

    2024

  6. [13]

    Exploration of llms, eeg, and behavioral data to measure and support attention and sleep.arXiv preprint arXiv:2408.07822, 2024

    Akane Sano, Judith Amores, and Mary Czerwinski. Exploration of llms, eeg, and behavioral data to measure and support attention and sleep.arXiv preprint arXiv:2408.07822, 2024. 1, 5.4

    2024 arXiv

  7. [14]

    Exploring chatgpt’s potential in ecg interpretation and outcome prediction in emergency department.The American Journal of Emergency Medicine, 88:7–11, 2025

    Arian Zaboli, Francesco Brigo, Marta Ziller, Magdalena Massar, Marta Parodi, Gabriele Magnarelli, Gloria Brigiari, and Gianni Turcato. Exploring chatgpt’s potential in ecg interpretation and outcome prediction in emergency department.The American Journal of Emergency Medicine,...

    2025

  8. [15]

    Springer, 2005

    Vladimir Vovk, Alexander Gammerman, and Glenn Shafer.Algorithmic learning in a random world, volume 29. Springer, 2005. 1, 2.3 24

    2005

  9. [16]

    Deep learning for ecg arrhythmia detection and classification: an overview of progress for period 2017–2023.Frontiers in Physiology, 14:1246746, 2023

    Yaqoob Ansari, Omar Mourad, Khalid Qaraqe, and Erchin Serpedin. Deep learning for ecg arrhythmia detection and classification: an overview of progress for period 2017–2023.Frontiers in Physiology, 14:1246746, 2023. 2.1

    2017

  10. [17]

    Seizure prediction—ready for a new era.Nature Reviews Neurology, 14(10):618–630, 2018

    Levin Kuhlmann, Klaus Lehnertz, Mark P Richardson, Bj¨ orn Schelter, and Hitten P Zaveri. Seizure prediction—ready for a new era.Nature Reviews Neurology, 14(10):618–630, 2018. 2.1

    2018

  11. [18]

    Research and application of deep learning-based sleep staging: Data, modeling, validation, and clinical practice.Sleep Medicine Reviews, page 101897, 2024

    Huijun Yue, Zhuqi Chen, Wenbin Guo, Lin Sun, Yidan Dai, Yiming Wang, Wenjun Ma, Xiaomao Fan, Weiping Wen, and Wenbin Lei. Research and application of deep learning-based sleep staging: Data, modeling, validation, and clinical practice.Sleep Medicine Reviews, page 101897, 2024. 2.1

    2024

  12. [19]

    Atrial activity extraction from single lead ecg recordings: Evaluation of two novel methods.Computers in biology and medicine, 43(3):176–183, 2013

    Huhe Dai, Shouda Jiang, and Ye Li. Atrial activity extraction from single lead ecg recordings: Evaluation of two novel methods.Computers in biology and medicine, 43(3):176–183, 2013. 2.1

    2013

  13. [20]

    Wavelet-based eeg processing for computer-aided seizure detection and epilepsy diagnosis.Seizure, 26:56–64, 2015

    Oliver Faust, U Rajendra Acharya, Hojjat Adeli, and Amir Adeli. Wavelet-based eeg processing for computer-aided seizure detection and epilepsy diagnosis.Seizure, 26:56–64, 2015. 2.1

    2015

  14. [21]

    Automatic ecg diagnosis using convolutional neural network

    Roberta Avanzato and Francesco Beritelli. Automatic ecg diagnosis using convolutional neural network. Electronics, 9(6):951, 2020. 2.1

    2020

  15. [22]

    Moderntcn: A modern pure convolution structure for general time series analysis

    Donghao Luo and Xue Wang. Moderntcn: A modern pure convolution structure for general time series analysis. InThe Twelfth International Conference on Learning Representations, 2024. 2.1

    2024

  16. [23]

    Contrast everything: A hierarchical contrastive framework for medical time-series.Advances in Neural Information Processing Systems, 36, 2024

    Yihe Wang, Yu Han, Haishuai Wang, and Xiang Zhang. Contrast everything: A hierarchical contrastive framework for medical time-series.Advances in Neural Information Processing Systems, 36, 2024. 2.1

    2024

  17. [24]

    Attention is all you need.Advances in Neural Information Processing Systems, 2017

    A Vaswani. Attention is all you need.Advances in Neural Information Processing Systems, 2017. 2.1, 3.2

    2017

  18. [25]

    An attention-based deep learning approach for sleep stage classification with single-channel eeg

    Emadeldeen Eldele, Zhenghua Chen, Chengyu Liu, Min Wu, Chee-Keong Kwoh, Xiaoli Li, and Cuntai Guan. An attention-based deep learning approach for sleep stage classification with single-channel eeg. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 29:809–818,...

    2021

  19. [26]

    Mts-lof: medical time- series representation learning via occlusion-invariant features.IEEE Journal of Biomedical and Health Informatics, 2024

    Huayu Li, Ana S Carreon-Rascon, Xiwen Chen, Geng Yuan, and Ao Li. Mts-lof: medical time- series representation learning via occlusion-invariant features.IEEE Journal of Biomedical and Health Informatics, 2024. 2.1

    2024

  20. [27]

    Solving the multiple instance problem with axis-parallel rectangles.Artificial intelligence, 89(1-2):31–71, 1997

    Thomas G Dietterich, Richard H Lathrop, and Tom´ as Lozano-P´ erez. Solving the multiple instance problem with axis-parallel rectangles.Artificial intelligence, 89(1-2):31–71, 1997. 2.2

    1997

  21. [28]

    A unified approach to interpreting model predictions.Advances in neural information processing systems, 30:4765–4774, 2017

    M Scott, Lee Su-In, et al. A unified approach to interpreting model predictions.Advances in neural information processing systems, 30:4765–4774, 2017. 2.2

    2017

  22. [29]

    ” why should i trust you?” explaining the predictions of any classifier

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. ” why should i trust you?” explaining the predictions of any classifier. InProceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135–1144, 2016. 2.2

    2016

  23. [30]

    A practical bayesian framework for backpropagation networks.Neural computation, 4(3):448–472, 1992

    David JC MacKay. A practical bayesian framework for backpropagation networks.Neural computation, 4(3):448–472, 1992. 2.3

    1992

  24. [31]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. Ininternational conference on machine learning, pages 1050–1059. PMLR,

  25. [32]

    Monty hall and optimized conformal prediction to improve decision-making with llms.arXiv preprint arXiv:2501.00555, 2024

    Harit Vishwakarma, Alan Mishler, Thomas Cook, Niccolo Dalmasso, Natraj Raman, and Sumitra Ganesh. Monty hall and optimized conformal prediction to improve decision-making with llms.arXiv preprint arXiv:2501.00555, 2024. 2.3, 5.3, 3

    2024 arXiv

  26. [33]

    Attention-based deep multiple instance learning

    Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In International conference on machine learning, pages 2127–2136. PMLR, 2018. 3.2, A, A1 25

    2018

  27. [34]

    Transmil: Transformer based correlated multiple instance learning for whole slide image classification.Advances in neural information processing systems, 34:2136–2147, 2021

    Zhuchen Shao, Hao Bian, Yang Chen, Yifeng Wang, Jian Zhang, Xiangyang Ji, et al. Transmil: Transformer based correlated multiple instance learning for whole slide image classification.Advances in neural information processing systems, 34:2136–2147, 2021. 3.2

    2021

  28. [35]

    Theory, analysis, and best practices for sigmoid self-attention.arXiv preprint arXiv:2409.04431, 2024

    Jason Ramapuram, Federico Danieli, Eeshan Dhekane, Floris Weers, Dan Busbridge, Pierre Ablin, Tatiana Likhomanenko, Jagrit Digani, Zijin Gu, Amitis Shidani, et al. Theory, analysis, and best practices for sigmoid self-attention.arXiv preprint arXiv:2409.04431, 2024. 3.2

    2024 arXiv

  29. [36]

    John Wiley & Sons, 1999

    Thomas M Cover.Elements of information theory. John Wiley & Sons, 1999. 3.2

    1999

  30. [37]

    Inductive confidence machines for regression

    Harris Papadopoulos, Kostas Proedrou, Volodya Vovk, and Alex Gammerman. Inductive confidence machines for regression. InMachine learning: ECML 2002: 13th European conference on machine learning Helsinki, Finland, August 19–23, 2002 proceedings 13, pages 345–356. Springer, 2002. 3.3

    2002

  31. [39]

    Conformal prediction beyond exchangeability.The Annals of Statistics, 51(2):816–845, 2023

    Rina Foygel Barber, Emmanuel J Candes, Aaditya Ramdas, and Ryan J Tibshirani. Conformal prediction beyond exchangeability.The Annals of Statistics, 51(2):816–845, 2023. 3.3, 3.3

    2023

  32. [40]

    Analysis of a sleep-dependent neuronal feedback loop: the slow-wave microcontinuity of the eeg.IEEE Transactions on Biomedical Engineering, 47(9):1185–1194, 2000

    Bob Kemp, Aeilko H Zwinderman, Bert Tuk, Hilbert AC Kamphuisen, and Josefien JL Oberye. Analysis of a sleep-dependent neuronal feedback loop: the slow-wave microcontinuity of the eeg.IEEE Transactions on Biomedical Engineering, 47(9):1185–1194, 2000. 5.2

    2000

  33. [41]

    Ptb-xl, a large publicly available electrocardiography dataset.Scientific data, 7(1):1–15, 2020

    Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Dieter Kreiseler, Fatima I Lunze, Wojciech Samek, and Tobias Schaeffter. Ptb-xl, a large publicly available electrocardiography dataset.Scientific data, 7(1):1–15, 2020. 5.2

    2020

  34. [42]

    Edward A. Wolpert. A Manual of Standardized Terminology, Techniques and Scoring System for Sleep Stages of Human Subjects.Archives of General Psychiatry, 20(2):246–247, 02 1969. 5.2

    1969

  35. [43]

    Richard B Berry, Rita Brooks, Charlene E Gamaldo, Susan M Harding, Carole Marcus, Bradley V Vaughn, et al. The aasm manual for the scoring of sleep and associated events.Rules, Terminology and Technical Specifications, Darien, Illinois, American Academy of Sleep Medicine, 176(...

    2012

  36. [44]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024. 5.2

    2024 arXiv

  37. [45]

    Mimo-vl technical report, 2025

    Xiaomi LLM-Core Team. Mimo-vl technical report, 2025. 5.2

    2025

  38. [46]

    Appropriateness of cardiovascular disease prevention recommendations obtained from a popular online chat-based artificial intelligence model.Jama, 329(10):842–844, 2023

    Ashish Sarraju, Dennis Bruemmer, Erik Van Iterson, Leslie Cho, Fatima Rodriguez, and Luke Laffin. Appropriateness of cardiovascular disease prevention recommendations obtained from a popular online chat-based artificial intelligence model.Jama, 329(10):842–844, 2023. 5.4

    2023

  39. [47]

    Maham Bilal, Yumna Jamil, Dua Rana, and Hussain Haider Shah. Enhancing awareness and self- diagnosis of obstructive sleep apnea using ai-powered chatbots: the role of chatgpt in revolutionizing healthcare.Annals of Biomedical Engineering, 52(2):136–138, 2024. 5.4

    2024

  40. [48]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022. 6

    2022

  41. [49]

    Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in Neural Information Processing Systems, 33:9459–9474, 2020

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K¨ uttler, Mike Lewis, Wen-tau Yih, Tim Rockt¨ aschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in Neural Information Processing Syst...

    2020

  42. [50]

    Jos´ e A Joglar, Mina K Chung, Anastasia L Armbruster, Emelia J Benjamin, Janice Y Chyou, Edmond M Cronin, Anita Deswal, Lee L Eckhardt, Zachary D Goldberger, Rakesh Gopinathannair, et al. 2023 acc/aha/accp/hrs guideline for the diagnosis and management of atrial fibrillation:...

    2023

  43. [51]

    Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning

    Bin Li, Yin Li, and Kevin W Eliceiri. Dual-stream multiple instance learning network for whole slide image classification with self-supervised contrastive learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318–14328, 2021. A, A1

    2021

  44. [52]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    I Loshchilov. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. A

    2017 arXiv

  45. [53]

    A gentle introduction to conformal prediction and distribution-free uncertainty quantification.arXiv preprint arXiv:2107.07511, 2021

    Anastasios N Angelopoulos and Stephen Bates. A gentle introduction to conformal prediction and distribution-free uncertainty quantification.arXiv preprint arXiv:2107.07511, 2021. A

    2021 arXiv

  46. [54]

    Tsmixer: An all-mlp architecture for time series forecasting.arXiv preprint arXiv:2303.06053, 2023

    Si-An Chen, Chun-Liang Li, Nate Yoder, Sercan O Arik, and Tomas Pfister. Tsmixer: An all-mlp architecture for time series forecasting.arXiv preprint arXiv:2303.06053, 2023. A1 27 A Supplementary Materials: Stagnation of SSMs and ConMIL Con- formal Coverage Baselines and experi...

    2023 arXiv

  47. [55]

    As shown in its reasoning ⟨think⟩ block, the model correctly identified the presence of ”relatively low-amplitude” and ”high-frequency activity” throughout the epoch

    MiMo-VL-7B Stand-Alone PerformanceWhen prompted to act as a ”highly skilled sleep medicine expert AI” and analyze the raw EEG waveform, MiMo-VL-7B initiated a detailed visual analysis. As shown in its reasoning ⟨think⟩ block, the model correctly identified the presence of ”rel...

  48. [56]

    Interviews

    MiMo-VL-7B Performance with ConMIL Support:The interaction changed fundamentally when the model was augmented by ConMIL. The prompt was more structured, providing the LLM with: •A constrained, high-confidence prediction set from conformal prediction:{N1, N2}. • Per-class inter...

  49. [57]

    The heatmap highlights regions where attention is focused on N1 criteria

    **N1 (N1 | Attention Overlay)**: This stage typically shows alpha activity decreasing with sleep onset, transitioning to theta with sleep spindles and K- complexes. The heatmap highlights regions where attention is focused on N1 criteria. However, the EEG waveform does not sho...

  50. [58]

    The heatmap indicates areas where N2 criteria are met

    **N2 (N2 | Attention Overlay)**: N2 is characterized by the presence of sleep spindles and K-complexes, along with a decrease in alpha activity and an increase in theta activity . The heatmap indicates areas where N2 criteria are met. The EEG waveform shows the typical feature...

  51. [59]

    First, I'll examine the frequency and amplitude characteristics: - The waveform shows high-frequency , low-amplitude activity - There are no prominent delta waves (slow waves) - No prominent K-complexes or sleep spindles are visible - No obvious REM-related theta activity

  52. [60]

    Comparing with standard sleep stage criteria: - Wake: Typically shows high-frequency , low-amplitude beta/alpha waves (matches what we see) - N1 (light sleep): Shows low-amplitude theta waves with occasional sleep spindles (not present) - N2 (deeper sleep): Shows K-complexes a...

  53. [61]

    The pattern is consistent with wakefulness, as it shows the high-frequency , low-amplitude activity typical of the awake state

  54. [62]

    Analyzing this 30-second Fpz-Cz EEG epoch: The waveform displays high-frequency , low-amplitude activity characteristic of the awake state

    The lack of sleep spindles, K-complexes, and delta waves rules out N1, N2, N3, and REM stages. Analyzing this 30-second Fpz-Cz EEG epoch: The waveform displays high-frequency , low-amplitude activity characteristic of the awake state. I can identify several key features:

  55. [63]

    Frequency pattern: The EEG shows predominantly high-frequency activity (beta/alpha range), which is typical of wakefulness

  56. [64]

    Amplitude characteristics: The waves maintain relatively low amplitude throughout the epoch, without the characteristic slow-wave patterns seen in deeper sleep stages

  57. [65]

    Absence of specific sleep markers: - No K-complexes (large slow waves with sharp negative deflections) - No sleep spindles (brief bursts of oscillatory activity) - No prominent delta waves (slow waves) - No theta activity associated with REM sleep

  58. [66]

    Temporal stability: The pattern remains consistent throughout the 30- second epoch without the transitional features seen in sleep stage transitions. When compared against the established criteria for sleep stages: - Wake: Matches the high-frequency , low-amplitude pattern - N...

  59. [67]

    **P Waves**: The P waves are upright in leads II, III, and aVF, which is normal for a right-sided lead

  60. [68]

    **QRS Complexes**: The QRS complexes appear narrow without any signs of bundle branch block or wide QRS that would suggest conduction disturbances

  61. [69]

    **ST Segments and T Waves**: There are no significant ST-segment elevations or depressions, nor are there any broadened Q waves or T-wave inversions indicative of myocardial infarction or hypertrophy

  62. [70]

    Normal ECG

    **Ratios**: The R/S ratio in V1 suggests left ventricular strain, but it is not definitive for either hypertrophy or any other pathology. Given these observations, the most likely diagnosis based on this ECG is "Normal ECG. " However, it is crucial to remember that a single EC...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.