Pith. sign in

REVIEW 4 major objections 4 minor 46 references

Continuous Cardiac Arrest Prediction in ICU using PPG Foundation Model

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

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

Pith's one-line read This paper claims that in-hospital cardiac arrest can be predicted from a single finger PPG signal alone, with an average AUROC of 0.79 over a 24-hour window and 0.82 one hour before onset.

desk verdict Genuine first for PPG-only IHCA prediction, but the unreported patient-level split is load-bearing for the 0.79 AUROC. read the letter →

arxiv 2502.08612 v1 pith:6RAWMGMH submitted 2025-02-12 cs.LG

Saurabh Kataria , Ran Xiao , Timothy Ruchti , Matthew Clark , Jiaying Lu , Randall J. Lee , Jocelyn Grunwell , Xiao Hu This is my paper · ORCID
classification cs.LG
keywords in-hospitalcardiacarrestphotoplethysmography(PPG)foundationmodelPPG-GPTICUmonitoringsequencemodelingAUROCprediction
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

This paper tries to establish that in-hospital cardiac arrest (IHCA) can be predicted from a single continuous finger photoplethysmography (PPG) signal, with no ECG, lab values, or other vital signs. The authors claim this is the first demonstration of IHCA prediction in ICU patients using only unimodal continuous PPG waveform representations. Their two-stage Feature Extractor-Aggregator Network (FEAN) pairs a pre-trained PPG foundation model (PPG-GPT) with a sequence aggregator, and their best model, a fine-tuned 345M-parameter extractor with an attention-bidirectional-LSTM aggregator, reaches an average area under the ROC curve (AUROC) of 0.79 over the 24 hours before cardiac arrest and 0.82 one hour before onset. If the result holds, a cheap finger sensor could deliver the kind of cardiac-arrest risk stratification that previously required multi-modal monitoring.

What carries the argument

Feature Extractor-Aggregator Network (FEAN): a two-stage cascade whose first stage is PPG-GPT, a pre-trained generative transformer that was trained to predict the next one-second patch of PPG and is used frozen or with light fine-tuning to embed each non-overlapping 30-second signal chunk (the last hidden state of the final layer becomes the chunk's embedding). The second stage is a sequence aggregator that combines these per-chunk embeddings into one vector for binary prediction; the default is BLSTM-Att, a bidirectional LSTM with learned attention, with Mamba and xLSTM tested as alternatives. The mechanism works by converting long PPG streams into a fixed-length health-state embedding and letting a sequential model capture how that state evolves over one to twenty-four hours.

What would settle it

A strict patient-level train/test split (no patient contributing segments to both sides) with the same FEAN-345M protocol would settle the claim: if the 24-hour average AUROC falls materially below 0.79, the published number reflects within-patient leakage rather than generalization to new patients.

Watch

Extended reading notes

Core claim

On the UCSF ICU dataset (200 cases, 1000 controls), the paper reports that fine-tuning the PPG-GPT feature extractor is what unlocks the gains: frozen one-hour models stay around 0.64-0.66 AUROC, while the tuned 345M model averages 0.79 over the 24-hour prediction window and reaches 0.82 at T-1. The full-history variant, evaluated only with a frozen extractor, reaches 0.76 AUROC and beats the frozen one-hour baseline, suggesting that longer patient history helps. Among aggregators, BLSTM with attention outperforms xLSTM and Mamba in this low-resource setting. The paper also visualizes PPG-GPT embeddings with PaCMAP and observes a consistent latent-space trajectory as arrest approaches, which it interprets as evidence that the model tracks health deterioration over time.

Load-bearing premise

The load-bearing assumption is that the one-hour training segments and the one-hour evaluation segments are independent at the patient level, because the paper does not describe a patient-level split; if the same 200 cases and 1000 controls contribute to both sets, within-patient similarity could inflate the reported 0.79 AUROC.

Editorial extensions

If this is right

  • A single finger PPG channel could replace multi-modal inputs for cardiac-arrest risk stratification in the ICU, simplifying monitoring and reducing alarm load.
  • Fine-tuning the foundation model is the key lever: on the one-hour model it lifts AUROC from roughly 0.64 to 0.79, so representation adaptation to the target ICU population matters more than model size alone.
  • Longer context helps when the extractor is frozen: the full-history variant reaches 0.76 AUROC versus 0.66 for the one-hour frozen baseline, implying that deterioration signals accumulate over hours.
  • If the latent-space trajectory is genuinely tied to patient state, the same embeddings could support continuous risk scores or trend alarms rather than binary predictions.

Reading between the lines

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

  • The paper leaves unstated whether the 0.79 figure would survive external validation on a different hospital's ICU data; that test is needed before treating finger PPG as a drop-in replacement for multi-modal monitoring.
  • The non-monotonic scaling with model size (tuned 345M beats tuned 1B) hints that on 200 positive cases the largest model overfits; patient-level splitting or stronger regularization might change the ranking.
  • Zero-filling missing or flat signal gaps is a choice that could inject artifacts; comparing zero-filling with imputation or masking would reveal how much of the predictive signal comes from the raw waveform itself.
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 / 4 minor

Summary. The paper proposes FEAN, a two-stage architecture that combines a pre-trained PPG foundation model (PPG-GPT, up to 1B parameters) with a sequential aggregator (BLSTM-Att, Mamba, or xLSTM) to predict in-hospital cardiac arrest (IHCA) from a single-channel finger PPG signal. Two variants are introduced: a 1-hour model (1H) using the latest hour of PPG, and a full-history model (FH) using up to 24 hours. On a UCSF ICU dataset with 200 case and 1000 control patients, the best 1H model (345M PPG-GPT with BLSTM-Att) achieves an average AUROC of 0.79 over the 24-hour prediction window, peaking at 0.82 at T-1. Baselines include hand-crafted morphological features and STFT features. The paper additionally provides architectural comparisons and PaCMAP trajectory visualizations.

Significance. If the reported AUROC survives a proper patient-level evaluation with confidence intervals, the paper would make an important contribution: it would show that a single non-invasive PPG channel can stratify cardiac arrest risk a day in advance at a level comparable to prior multimodal models, and it would demonstrate the utility of large PPG foundation models in a low-resource clinical prediction task. The study uses an externally adjudicated ICU dataset (UCSF), includes conventional baselines (morphological and STFT features), provides a public code repository link, and reports raw hourly AUROC values in the appendix. These are concrete strengths. However, the evaluation protocol as described does not yet support the headline claim: no patient-level train/test split is reported, no confidence intervals are given, and the zero-padding policy may allow a missingness shortcut. The central claim therefore remains unvalidated until these evaluative gaps are addressed.

major comments (4)
  1. [Section II.C] The manuscript never describes a patient-level split of the 200 case and 1000 control patients into training and test sets. Section II.C states that during training a random 1-hour subsequence is sampled from [T-25, T-1] for each patient, and that during evaluation predictions are made hourly over the same window. If the same patients contribute segments to both training and evaluation, the model can memorize patient-specific waveform characteristics, and the 24 hourly predictions per patient are highly autocorrelated; the averaged AUROC of 0.79 would then reflect intra-patient similarity rather than generalization to new patients. Please specify whether any patient appeared in both training and testing, report the number of distinct patients in each fold, and provide metrics computed with patient-level clustering or a patient-disjoint split.
  2. [Section II.B] The zero-padding rule ('For time gaps when the signal is absent or flat, we simply insert zeroes') creates a potential shortcut: if missing or flat intervals are more common in case patients before arrest, the model could learn to classify missingness rather than physiological deterioration. This would inflate AUROC without providing clinically meaningful prediction. Please report the amount and distribution of zero-padded (or flat) intervals in cases versus controls, and provide an ablation that controls for missingness—for example, adding a missingness indicator as a feature, masking padded segments, or evaluating only on segments with sufficient signal quality.
  3. [Table I and Section III.B] The text states that 'the performance correlates with the size of the foundation model,' but Table I shows GPT-345M at AUROC 0.791 and GPT-1B at 0.749, with GPT-85M at 0.786. This is not a monotonic relationship and the claim is contradicted by the presented numbers. The conclusion should be revised to describe performance peaking at 345M and degrading at 1B, or the authors should explain any difference in tuning. In addition, all AUROC values are point estimates without confidence intervals; given the small case count (200 patients), the difference between 0.791 and 0.749 may not be significant.
  4. [Appendix Table IV] The caption of Table IV contains an incomplete sentence: 'The lack of consistency in patterns is' ends without a predicate. This is not just a typographical issue; the table itself shows large hour-to-hour fluctuations (e.g., STFT AUROC ranges from 0.508 at T-7 to 0.721 at T-12), so any claim about trends or averages must be accompanied by variance estimates. Please complete the caption and add confidence intervals or error bars for the hourly AUROC values.
minor comments (4)
  1. [Section II.A.1] The description of PPG-GPT feature extraction says 'we use the last timestamp from the hidden layer output of the last layer' and that inputs longer than 30 s are split into 30 s chunks processed independently. It would be clearer to specify how the embeddings from multiple 30 s chunks are aligned for the aggregator (e.g., chunk order and whether overlapping windows are used).
  2. [Section III.A] The PaCMAP visualization in Fig. 3 is described as 'potentially capturing the patient's health trajectory in the latent space.' This is a qualitative illustration based on two patients; the caption and text would benefit from stating explicitly that no quantitative evaluation of trajectory separability is provided.
  3. [Section III.C] The sentence 'Due to low-resource data availability, we notice a high variance of our model performance w.r.t. the choice of hyperparameters' is not backed by any reported variance or hyperparameter sensitivity results. Please provide the observed ranges or refer to a supplementary experiment.
  4. [Section II.B] The case-control ratio is 1:5 (200/1000), but the appendix Fig. 5 states 'prevalence is 11%.' These numbers are inconsistent unless the evaluation set is rebalanced; clarify how prevalence is computed, because the AUPRC values in Tables I–III depend on the actual class ratio used in evaluation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the AUROC is computed against external clinical adjudication labels; self-cited PPG-GPT is an executed artifact, and no result is defined in terms of its own inputs.

full rationale

The derivation chain is: raw single-channel PPG to PPG-GPT (pretrained next-patch prediction) to FEAN aggregator to binary cross-entropy against clinically adjudicated cardiac arrest onset times. The reported AUROC is a function of those external labels, not of the pretraining objective or of any quantity that contains the prediction itself. The PPG-GPT citation [4] and dataset citations [19, 32] involve overlapping authors, but they supply a model artifact and a data source that are actually executed and evaluated; the pretraining target (next-patch prediction on 30-s PPG chunks) does not include the cardiac-arrest label, so the downstream result is not forced by the self-citation. Independent baselines (morphological features and STFT) are reported and are worse, providing a non-circular yardstick. Two methodological concerns are noted but are not circularity: Section II.C never documents a patient-level train/test split (training samples random 1-hour subsequences from [T-25, T-1] while evaluation makes hourly predictions over that same window), which is an external-validity/leakage risk; and Section II.B's zero-padding of absent/flat intervals could let the model exploit missingness, which is a confounding-signal risk. Neither concern makes the reported AUROC equivalent by definition to the model's training inputs, so no circular step can be exhibited from the text.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard supervised learning assumptions plus a few domain-specific choices. The most fragile are the zero-padding rule for missing signal and the unstated patient-level independence of train and test samples.

free parameters (4)
  • Prediction horizon = 24 hours
    The label is 'cardiac arrest within the next 24 hours' and the headline AUROC is averaged over hourly predictions in this window. The choice is arbitrary and affects the metric.
  • Lead time = 1 hour
    The last hour before arrest is excluded from training and evaluation inputs, which avoids modeling the resuscitation period but introduces a lead-time choice.
  • Case-control sampling = 200 cases, 1000 controls
    The 1:5 imbalance sets the baseline AUPRC (0.106) and influences precision-recall metrics.
  • Alarm frequency = 1 hour
    Predictions are emitted every hour, giving 24 evaluation points per patient. A different frequency would change the averaged AUROC.
assumptions (4)
  • domain assumption Zero-padding of absent or flat PPG segments is a faithful representation of missing data.
    Section II.B states 'For time gaps when the signal is absent or flat, we simply insert zeroes.' This assumes missingness is not itself predictive.
  • domain assumption Training and evaluation examples are independent at the patient level.
    The paper does not describe a patient-level split, so the reported evaluation implicitly relies on this assumption.
  • domain assumption The UCSF dataset contains rigorous clinical adjudication of cardiac arrest events.
    Section II.B describes the dataset as having 'rigorous clinical adjudication of cardiac arrest events'.
  • domain assumption PPG-GPT pretrained representations transfer to ICU PPG signals.
    The feature extractor was pretrained on UCSF data [4]; the paper assumes its representations are useful for this downstream task.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Continuous Cardiac Arrest Prediction in ICU using PPG Foundation Model." pith.science (2026). https://pith.science/paper/6RAWMGMH

@misc{pith2026250208612,
  author       = {Pith},
  title        = {Pith review of: Continuous Cardiac Arrest Prediction in ICU using PPG Foundation Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6RAWMGMH}},
  note         = {Machine review of arXiv:2502.08612}
}
read the original abstract

Non-invasive patient monitoring for tracking and predicting adverse acute health events is an emerging area of research. We pursue in-hospital cardiac arrest (IHCA) prediction using only single-channel finger photoplethysmography (PPG) signals. Our proposed two-stage model Feature Extractor-Aggregator Network (FEAN) leverages powerful representations from pre-trained PPG foundation models (PPG-GPT of size up to 1 Billion) stacked with sequential classification models. We propose two FEAN variants ("1H", "FH") which use the latest one-hour and (max) 24-hour history to make decisions respectively. Our study is the first to present IHCA prediction results in ICU patients using only unimodal (continuous PPG signal) waveform deep representations. With our best model, we obtain an average of 0.79 AUROC over 24~h prediction window before CA event onset with our model peaking performance at 0.82 one hour before CA. We also provide a comprehensive analysis of our model through architectural tuning and PaCMAP visualization of patient health trajectory in latent space.

Figures

Figures reproduced from arXiv: 2502.08612 by the authors.

Figure 1
Figure 1. Illustation of Feature Extractor-Aggregator Network [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the overall pipeline (for a case patient). 1-hour model (“1H”) and Full-history (“FH”) model are trained [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Illustration of moving average of PaCMAP mapping of PPG-GPT embeddings of 1 hour PPG data before cardiac [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Hourly AUROC for STFT baseline and 1H 345M (both feature extractor tuned and frozen). For the feature extractor [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Receiver Operating Curve (ROC) values averaged [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 38 canonical work pages

  1. [1]

    On the analysis of fingertip photoplethysmogram signals

    Mohamed Elgendi. On the analysis of fingertip photoplethysmogram signals. Current cardiology reviews, 8(1):14–25, 2012

    work page 2012

  2. [2]

    Wrist photoplethysmography signal quality assessment for reliable heart rate estimate and morphological analysis

    Serena Moscato, Stella Lo Giudice, Giulia Massaro, and Lorenzo Chiari. Wrist photoplethysmography signal quality assessment for reliable heart rate estimate and morphological analysis. Sensors, 22(15):5831, 2022. TABLE IV: Raw AUROC values obtained for three models (STFT Baseline, 345M Frozen, 345M Tuned). The lack of consistency in patterns is Time of pr...

    work page 2022

  3. [3]

    Highly wearable cuff- less blood pressure and heart rate monitoring with single-arm electro- cardiogram and photoplethysmogram signals

    Qingxue Zhang, Dian Zhou, and Xuan Zeng. Highly wearable cuff- less blood pressure and heart rate monitoring with single-arm electro- cardiogram and photoplethysmogram signals. Biomedical engineering online, 16:1–20, 2017

    work page 2017

  4. [4]

    Adapting a generative pretrained transformer achieves sota performance in assessing diverse physiological functions using only photoplethysmography signals: A gpt-ppg approach

    Zhaoliang Chen, Cheng Ding, Nirbhay Modhe, Jiaying Lu, Carl Yang, and Xiao Hu. Adapting a generative pretrained transformer achieves sota performance in assessing diverse physiological functions using only photoplethysmography signals: A gpt-ppg approach. In AAAI 2024 Spring Symposium on Clinical Foundation Models , 2024

    work page 2024

  5. [5]

    Papagei: Open foundation models for optical physio- logical signals

    Arvind Pillai, Dimitris Spathis, Fahim Kawsar, and Mohammad Malekzadeh. Papagei: Open foundation models for optical physio- logical signals. arXiv preprint arXiv:2410.20542 , 2024

    arXiv 2024

  6. [6]

    Large-scale train- ing of foundation models for wearable biosignals

    Salar Abbaspourazad, Oussama Elachqar, Andrew C Miller, Saba Emrani, Udhyakumar Nallasamy, and Ian Shapiro. Large-scale train- ing of foundation models for wearable biosignals. arXiv preprint arXiv:2312.05409, 2023

    arXiv 2023

  7. [7]

    Signal quality assessment of ppg signals using stft time-frequency spectra and deep learning approaches

    Jianzhong Chen, Ke Sun, Yi Sun, and Xinxin Li. Signal quality assessment of ppg signals using stft time-frequency spectra and deep learning approaches. In 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) , pages 1153–1156. IEEE, 2021

    work page 2021

  8. [8]

    Detecting cardiac states with wearable photoplethysmograms and implications for out-of-hospital cardiac arrest detection

    Mahsa Khalili, Saud Lingawi, Jacob Hutton, Christopher B Fordyce, Jim Christenson, Babak Shadgan, Brian Grunau, and Calvin Kuo. Detecting cardiac states with wearable photoplethysmograms and implications for out-of-hospital cardiac arrest detection. Scientific Reports, 14(1):23185, 2024

    work page 2024

Show all 46 references
  1. [9]

    Estimation of aortic stiffness by finger photoplethys- mography using enhanced pulse wave analysis and machine learning

    Henrik Hellqvist, Mikael Karlsson, Johan Hoffman, Thomas Kahan, and Jonas Spaak. Estimation of aortic stiffness by finger photoplethys- mography using enhanced pulse wave analysis and machine learning. Frontiers in Cardiovascular Medicine , 11:1350726, 2024

    2024

  2. [10]

    Biometric au- thentication using the ppg: A long-term feasibility study

    Jorge Sancho, ´Alvaro Alesanco, and Jos ´e Garc ´ıa. Biometric au- thentication using the ppg: A long-term feasibility study. Sensors, 18(5):1525, 2018

    2018

  3. [11]

    A review of heartbeat detection systems for automotive applications

    Toshiya Arakawa. A review of heartbeat detection systems for automotive applications. Sensors, 21(18):6112, 2021

    2021

  4. [12]

    Rehabilitation assessment system for stroke patients based on fusion-type optoelectronic plethysmography device and multi- modality fusion model: Design and validation

    Liangwen Yan, Ze Long, Jie Qian, Jianhua Lin, Sheng Quan Xie, and Bo Sheng. Rehabilitation assessment system for stroke patients based on fusion-type optoelectronic plethysmography device and multi- modality fusion model: Design and validation. Sensors, 24(9):2925, 2024

    2024

  5. [13]

    Yong Bai, Duc Do, Quan Ding, Jorge Arroyo Palacios, Yalda Shahriari, Michele M Pelter, Noel Boyle, Richard Fidler, and Xiao Hu. Is the sequence of superalarm triggers more predictive than sequence of the currently utilized patient monitor alarms? IEEE Transactions on Biomedica...

    2016

  6. [14]

    Usefulness of trends in continuous elec- trocardiographic telemetry monitoring to predict in-hospital cardiac arrest

    Duc H Do, Alan Kuo, Edward S Lee, David Mortara, David Elashoff, Xiao Hu, and Noel G Boyle. Usefulness of trends in continuous elec- trocardiographic telemetry monitoring to predict in-hospital cardiac arrest. The American journal of cardiology , 124(7):1149–1158, 2019

    2019

  7. [15]

    A preliminary study of continuous signals in predictive modeling of cardiac arrest in pediatric icu pa- tients using electrocardiography and photoplethysmography analytics

    Yujie Yan, Sukrit Treewaree, Jiahui Yao, Jiwoo Noh, Sheel Tanna, Joseph L Greenstein, Casey O Taylor, Timothy Ruchti, Jules Bergmann, and James Fackler. A preliminary study of continuous signals in predictive modeling of cardiac arrest in pediatric icu pa- tients using electro...

    2024

  8. [16]

    Early detection of sepsis utilizing deep learning on electronic health record event sequences

    Simon Meyer Lauritsen, Mads Ellersgaard Kalør, Emil Lund Kongs- gaard, Katrine Meyer Lauritsen, Marianne Johansson Jørgensen, Jeppe Lange, and Bo Thiesson. Early detection of sepsis utilizing deep learning on electronic health record event sequences. Artificial Intelligence in...

    2020

  9. [17]

    Automated myocardial infarction and angina detection using second derivative of photoplethysmography

    Neha, HK Sardana, N Dahiya, N Dogra, R Kanawade, YP Sharma, and S Kumar. Automated myocardial infarction and angina detection using second derivative of photoplethysmography. Physical and Engineering Sciences in Medicine , 46(3):1259–1269, 2023

    2023

  10. [18]

    An algorithm strategy for precise patient monitoring in a connected healthcare enterprise

    Xiao Hu. An algorithm strategy for precise patient monitoring in a connected healthcare enterprise. NPJ Digital Medicine, 2(1):30, 2019

    2019

  11. [19]

    Insights into the problem of alarm fatigue with physiologic monitor devices: a comprehensive observational study of consecutive intensive care unit patients

    Barbara J Drew, Patricia Harris, Jessica K Z `egre-Hemsey, Tina Mammone, Daniel Schindler, Rebeca Salas-Boni, Yong Bai, Adelita Tinoco, Quan Ding, and Xiao Hu. Insights into the problem of alarm fatigue with physiologic monitor devices: a comprehensive observational study of c...

    2014

  12. [20]

    Predictive combinations of monitor alarms preceding in-hospital code blue events

    Xiao Hu, Monica Sapo, Val Nenov, Tod Barry, Sunghan Kim, Duc H Do, Noel Boyle, and Neil Martin. Predictive combinations of monitor alarms preceding in-hospital code blue events. Journal of biomedical informatics, 45(5):913–921, 2012

    2012

  13. [21]

    Validating the electronic cardiac arrest risk triage (ecart) score for risk stratification of surgical inpatients in the postoperative setting: retrospective cohort study

    Bartlomiej Bartkowiak, Ashley M Snyder, Andrew Benjamin, Andrew Schneider, Nicole M Twu, Matthew M Churpek, Kevin K Roggin, and Dana P Edelson. Validating the electronic cardiac arrest risk triage (ecart) score for risk stratification of surgical inpatients in the postoperativ...

    2019

  14. [22]

    Modified early warning score vs cardiac arrest risk triage score for prediction of cardiopulmonary arrest: A case–control study

    Armand Delo Antone Tan, Chito Caimoy Permejo, and Ma Consola- cion Dolor Torres. Modified early warning score vs cardiac arrest risk triage score for prediction of cardiopulmonary arrest: A case–control study. Indian Journal of Critical Care Medicine: Peer-reviewed, Official P...

    2022

  15. [23]

    Steven C Brooks, Gareth R Clegg, Janet Bray, Charles D Deakin, Gavin D Perkins, Mattias Ringh, Christopher M Smith, Mark S Link, Raina M Merchant, Jaime Pezo-Morales, et al. Optimizing outcomes after out-of-hospital cardiac arrest with innovative approaches to public-access de...

    2022

  16. [24]

    Artificial intelligence algorithm for predicting cardiac arrest using electrocardiography

    Joon-myoung Kwon, Kyung-Hee Kim, Ki-Hyun Jeon, Soo Youn Lee, Jinsik Park, and Byung-Hee Oh. Artificial intelligence algorithm for predicting cardiac arrest using electrocardiography. Scandinavian journal of trauma, resuscitation and emergency medicine , 28:1–10, 2020

    2020

  17. [25]

    Real-time machine learning model to predict in-hospital cardiac arrest using heart rate variability in icu

    Hyeonhoon Lee, Hyun-Lim Yang, Ho Geol Ryu, Chul-Woo Jung, Youn Joung Cho, Soo Bin Yoon, Hyun-Kyu Yoon, and Hyung-Chul Lee. Real-time machine learning model to predict in-hospital cardiac arrest using heart rate variability in icu. NPJ Digital Medicine , 6(1):215, 2023

    2023

  18. [26]

    Prediction of cardiac arrest in critically ill patients based on bedside vital signs monitoring

    Li Yijing, Ye Wenyu, Yang Kang, Zhang Shengyu, He Xianliang, Jin Xingliang, Wang Cheng, Sun Zehui, and Liu Mengxing. Prediction of cardiac arrest in critically ill patients based on bedside vital signs monitoring. Computer Methods and Programs in Biomedicine , 214:106568, 2022

    2022

  19. [27]

    Prediction of in-hospital cardiac arrest in the intensive care unit: Machine learning–based multimodal approach

    Hsin-Ying Lee, Po-Chih Kuo, Frank Qian, Chien-Hung Li, Jiun-Ruey Hu, Wan-Ting Hsu, Hong-Jie Jhou, Po-Huang Chen, Cho-Hao Lee, Chin-Hua Su, et al. Prediction of in-hospital cardiac arrest in the intensive care unit: Machine learning–based multimodal approach. JMIR Medical Infor...

    2024

  20. [28]

    Abstract 4136643: Automated cardiac arrest detection using a photoplethysmography wristband: Evaluation of the time to detection of the circulatory arrest event

    Roos Edgar, Niels Scholte, Kambiz Ebrahimkheil, Marc Brouwer, Rypko Beukema, Masih Mafi-Rad, Kevin Vernooy, Sing Yap, Eelko Ronner, Nicolas Van Mieghem, Eric Boersma, Aysun Cetinyurek- Yavuz, Peter Stas, Niels Royen, and Judith Bonnes. Abstract 4136643: Automated cardiac arres...

    2024

  21. [29]

    Assessment of vasoactive agents and vascular aging by the second derivative of photoplethysmogram waveform

    Kenji Takazawa, Nobuhiro Tanaka, Masami Fujita, Osamu Mat- suoka, Tokuyu Saiki, Masaru Aikawa, Sinobu Tamura, and Chiharu Ibukiyama. Assessment of vasoactive agents and vascular aging by the second derivative of photoplethysmogram waveform. Hypertension, 32(2):365–370, 1998

    1998

  22. [30]

    Features from the pho- toplethysmogram and the electrocardiogram for estimating changes in blood pressure

    Eoin Finnegan, Shaun Davidson, Mirae Harford, Peter Watkinson, Lionel Tarassenko, and Mauricio Villarroel. Features from the pho- toplethysmogram and the electrocardiogram for estimating changes in blood pressure. Scientific Reports, 13(1):986, 2023

    2023

  23. [31]

    Foundation models for physiological data, how to develop them, and what to expect of them

    Xiao Hu. Foundation models for physiological data, how to develop them, and what to expect of them. Physiological Measurement , 45(2):020301, 2024

    2024

  24. [32]

    Siamquality: a convnet-based foundation model for photoplethysmography signals

    Cheng Ding, Zhicheng Guo, Zhaoliang Chen, Randall J Lee, Cynthia Rudin, and Xiao Hu. Siamquality: a convnet-based foundation model for photoplethysmography signals. Physiological Measurement, 45(8):085004, aug 2024

    2024

  25. [33]

    Universal pooling method of multi-layer features from pretrained models for speaker verification

    Jin Sob Kim, Hyun Joon Park, Wooseok Shin, and Sung Won Han. Universal pooling method of multi-layer features from pretrained models for speaker verification. arXiv preprint arXiv:2409.07770 , 2024

    2024

  26. [34]

    Attention is all you need

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

    2017

  27. [35]

    Flashattention: Fast and memory-efficient exact attention with io- awareness

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher R ´e. Flashattention: Fast and memory-efficient exact attention with io- awareness. Advances in Neural Information Processing Systems , 35:16344–16359, 2022

    2022

  28. [36]

    Long short-term memory

    J ¨urgen Schmidhuber, Sepp Hochreiter, et al. Long short-term memory. Neural Comput, 9(8):1735–1780, 1997

    1997

  29. [37]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752 , 2023

    2023 arXiv

  30. [38]

    xlstm: Extended long short-term memory

    Maximilian Beck, Korbinian P ¨oppel, Markus Spanring, Andreas Auer, Oleksandra Prudnikova, Michael Kopp, G ¨unter Klambauer, Johannes Brandstetter, and Sepp Hochreiter. xlstm: Extended long short-term memory. arXiv preprint arXiv:2405.04517 , 2024

    2024 arXiv

  31. [39]

    Mimic-iv, a freely accessible electronic health record dataset

    Alistair EW Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Ayad Shammout, Steven Horng, Tom J Pollard, Sicheng Hao, Benjamin Moody, Brian Gow, et al. Mimic-iv, a freely accessible electronic health record dataset. Scientific data, 10(1):1, 2023

    2023

  32. [40]

    Understanding how dimension reduction tools work: an empirical approach to deciphering t-sne, umap, trimap, and pacmap for data visualization

    Yingfan Wang, Haiyang Huang, Cynthia Rudin, and Yaron Shaposh- nik. Understanding how dimension reduction tools work: an empirical approach to deciphering t-sne, umap, trimap, and pacmap for data visualization. Journal of Machine Learning Research , 22(201):1–73, 2021

    2021

  33. [41]

    pyppg: A python toolbox for comprehensive photoplethysmography signal analysis

    M ´arton ´A Goda, Peter H Charlton, and Joachim A Behar. pyppg: A python toolbox for comprehensive photoplethysmography signal analysis. Physiological Measurement, 45(4):045001, 2024

    2024

  34. [42]

    Mars: Unleashing the power of variance reduction for training large models

    Huizhuo Yuan, Yifeng Liu, Shuang Wu, Xun Zhou, and Quanquan Gu. Mars: Unleashing the power of variance reduction for training large models. arXiv preprint arXiv:2411.10438 , 2024

    2024 arXiv

  35. [43]

    Development of a real-time risk prediction model for in-hospital cardiac arrest in critically ill patients using deep learning: retrospective study

    Junetae Kim, Yu Rang Park, Jeong Hoon Lee, Jae-Ho Lee, Young-Hak Kim, Jin Won Huh, et al. Development of a real-time risk prediction model for in-hospital cardiac arrest in critically ill patients using deep learning: retrospective study. JMIR Medical Informatics, 8(3):e16349, 2020

    2020

  36. [44]

    Cumulative time series representation for code blue prediction in the intensive care unit

    Rebeca Salas-Boni, Yong Bai, and Xiao Hu. Cumulative time series representation for code blue prediction in the intensive care unit. AMIA Summits on Translational Science Proceedings , 2015:162, 2015

    2015

  37. [45]

    Explainable artificial intelligence warning model using an ensemble approach for in-hospital cardiac arrest prediction: Retrospective cohort study

    Yun Kwan Kim, Ja Hyung Koo, Sun Jung Lee, Hee Seok Song, and Minji Lee. Explainable artificial intelligence warning model using an ensemble approach for in-hospital cardiac arrest prediction: Retrospective cohort study. Journal of Medical Internet Research , 25:e48244, 2023

    2023

  38. [46]

    Generalizability of superalarm via cross-institutional performance evaluation

    Ran Xiao, Duc Do, Cheng Ding, Karl Meisel, Randall J Lee, and Xiao Hu. Generalizability of superalarm via cross-institutional performance evaluation. IEEE Access, 8:132404–132412, 2020

    2020

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.