Pith. sign in

REVIEW 3 major objections 5 minor 38 references

Early Risk Prediction of Pediatric Cardiac Arrest from Electronic Health Records via Multimodal Fused Transformer

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read A multimodal fused transformer that reads the first 24 hours of EHR data as both numbers and text can predict pediatric cardiac arrest and beats ten comparison models.

desk verdict A competent multimodal fusion application to pediatric cardiac arrest, but the missing exclusion of early-arrest admissions and the overstated 'five metrics' claim need fixing before the empirical result is trusted. read the letter →

arxiv 2502.07158 v3 pith:OIXEKWBB submitted 2025-02-11 cs.LG cs.AI

classification cs.LGcs.AI
keywords pediatriccardiacarrestearlyriskpredictionelectronichealthrecordsmultimodalfusiontransformertabulardatatextualizationintensivecareunit
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 tries to establish that a multimodal transformer, PEDCA-FT, can predict pediatric cardiac arrest from the first 24 hours of electronic health record (EHR) data in a cardiac ICU, and that it does so better than ten comparison models. The reason this matters is that arrests in children are often preceded by a period of instability, so a reliable early warning from already-recorded data could give clinicians time to intervene. The paper reports the highest balanced accuracy (63.08%), F1 (14.31%), Matthews correlation coefficient (13.18%), and area under the precision-recall curve (9.15%) among all compared models on a cohort of 3,566 pediatric patients with a 4.0% arrest rate, while ranking second on AUROC (73.99%). It also claims that the features the model relies on, including capillary refill time, bilirubin, hematocrit, CO2, and FiO2, match clinically recognized warning signs.

What carries the argument

The load-bearing mechanism is the textualization function $g(\cdot)$, which rewrites every EHR factor as readable text: static factors become 'name: value', categorical time series become the set of distinct values, numerical time series become their min-max range, and normal-range lab results are filtered out so only abnormal findings remain, with section headers for demographics, vitals, assessments, labs, and medications. This derived textual view is fed to a pretrained bidirectional text transformer, while the raw tabular view, consisting of static factors plus the last observed value of each temporal factor, is fed to a tabular transformer. A late-fusion transformer combines the two view-specific embeddings and outputs the arrest probability, trained with focal loss to counter the roughly 3-4% arrest rate. The mechanism's work is to give the text encoder access to dynamics and value ranges that the tabular aggregation discards, while the tabular encoder retains the high-dimensional static structure.

What would settle it

Restrict the same cohort to admissions in which cardiac arrest occurred after hour 24 and rerun the five-fold comparison; if balanced accuracy or AUPRC drops substantially, the reported gains come from post-arrest measurements inside the feature window rather than from early prediction.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that treating the first 24 hours of a pediatric cardiac ICU admission as two complementary views, a tabular view of static demographics plus last-observed values of 184 temporal risk factors and a textual view that renders each factor as a phrase such as 'Sex: Female' or 'Body Temp: 91.4-98.4' while listing only abnormal lab values under section headers, lets a late-fusion transformer extract more predictive signal than either view alone. The fused model PEDCA-FT reports balanced accuracy 63.08%, F1 14.31%, MCC 13.18%, AUPRC 9.15%, and AUROC 73.99% under five-fold cross-validation, outperforming all ten baselines on the first four metrics and trailing only the best time-series model on AUROC. The paper reads this comparison as evidence that multimodal fusion captures dynamics that tabular aggregation loses and high-dimensional structure that pure time-series models miss, and the permutation-importance analysis is offered as support that the learned risk factors are clinically meaningful.

Load-bearing premise

The paper defines the prediction target as arrest during the rest of the admission after the first 24 hours, but it never states that patients who arrest within those first 24 hours are excluded from the cohort, so the model's input window could contain measurements taken after the arrest and thereby leak the label.

Editorial extensions

If this is right

  • A cardiac ICU could compute an arrest-risk probability within the first 24 hours from data already in the EHR, without extra charting or proprietary scoring tools.
  • The comparison supports multimodal fusion over single-view modeling, since the fused model leads on four of five metrics while tabular baselines lead on balanced accuracy among the baseline groups and the best time-series baseline leads on AUROC.
  • If the permutation-importance ranking holds, capillary refill time, total bilirubin, hematocrit, arterial CO2, and FiO2 are the first-day measurements that deserve the closest monitoring.
  • The AUROC-versus-balanced-metrics pattern shows that evaluation choices matter in imbalanced arrest screening: a model can rank patients well overall yet miss the high-risk minority that a clinical early-warning system needs.

Reading between the lines

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

  • A natural external-validation test is to apply the same textualization template to another pediatric ICU database without retuning; if the performance transfers, the representation itself, rather than the specific cohort, is driving the result.
  • The unresolved label-construction question can be settled by re-running the cohort with first-24-hour arrests excluded; until that check is published, the reported margins over the best tabular and time-series baselines should be read with that caveat.
  • Because the textual view discards normal-range lab values, the model's probabilities are tied to the reference ranges used at the source site, so cross-hospital deployment may require local normal-range tables.
  • Decision-curve or net-benefit analysis, rather than AUROC alone, would clarify whether the balanced-metric improvements translate into fewer missed arrests at a tolerable false-alarm rate.
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

3 major / 5 minor

Summary. The paper introduces PedCA-FT, a late-fusion transformer that combines a tabular transformer over static and last-observed EHR features with a pre-trained BERT-style text transformer over a textualized representation (value ranges, distinct categorical values, abnormal labs) of the first 24 hours of admission. It is evaluated on a private CHOA-CICU cohort of 3,566 pediatric patients (4,672 admissions, 3.1% CA incidence) using 5-fold cross-validation with patient-level splitting. The model is compared against ten baselines from three families (PEWS-based time-series models, last-observation tabular models, and numerical time-series deep networks) on balanced accuracy, F1, MCC, AUPRC, and AUROC, and a permutation feature importance analysis identifies features such as capillary refill, bilirubin, and hematocrit. The central claim is that PedCA-FT outperforms all ten models across five metrics and that multimodal fusion is beneficial.

Significance. If the empirical claims hold, PedCA-FT would be a clinically plausible early-warning tool for pediatric cardiac arrest in a CICU setting, and the textualization of structured EHR data is an interesting and potentially reusable technique for other risk-prediction tasks. The paper has concrete strengths: a clearly formulated prediction problem, a patient-level cross-validation split, a broad set of baseline models, use of focal loss to address class imbalance, and a feature importance analysis with clinical interpretation. The main value, however, rests on the integrity of the outcome timeline and on the statistical support for the claimed superiority; both currently need substantial clarification before the results can be accepted.

major comments (3)
  1. [Abstract and Section IV-C vs. Table II] The task definition predicts cardiac arrest 'during the rest of the admission stay' from data in the first τ = 24 hours, but Section II-B does not state that patients whose CA onset occurs before hour 24 are excluded from the positive class, nor how admissions shorter than 24 hours are handled. If such patients are included, the 24-hour feature window can contain measurements taken during or after resuscitation (e.g., post-arrest vitals, labs, medications), so the model would be reading the label from the input rather than performing early prediction. This would invalidate all metrics in Table II as estimates of early risk. Please state the exclusion criterion explicitly, report event-time statistics (e.g., the distribution of time from admission to CA onset), and, if needed, rerun the evaluation on a cohort where the feature window provably precedes the event.
  2. [Section IV-B and IV-C] The abstract claims that PedCA-FT 'outperforms ten other artificial intelligence models across five key performance metrics,' but Table II shows that gMLP achieves a higher AUROC (75.04 ± 3.52 vs. 73.99 ± 3.16). The text later acknowledges that PedCA-FT 'ranks second in AUROC,' so the abstract's claim is internally inconsistent and must be corrected. Furthermore, the statement that PedCA-FT 'marginally outperforms all compared models' is not supported by any significance testing: error bars overlap on AUPRC and AUROC (and partly on F1 and MCC), and no paired tests, confidence intervals, or fold-level comparisons are reported. Please add a statistical comparison across the cross-validation folds (e.g., paired bootstrap or signed-rank tests) and adjust the claims in the abstract and introduction accordingly.
  3. [Section IV-C] The comparison does not isolate the contribution of multimodal fusion: the tabular baselines use LAST aggregation, the time-series baselines use hourly mean aggregation, while the proposed textual view includes min-max ranges and all distinct categorical values, which provides strictly richer information. The reported gains might therefore be attributable to the more expressive input representation rather than to the fusion architecture itself. To support the claim that 'integrating a tabular transformer ... and a pre-trained textual transformer ... delivers SOTA performance' and that multimodal fusion is effective, the paper should include ablations: PedCA-FT with only the tabular encoder, only the textual encoder, and the full fusion model, all trained under the same preprocessing conditions.
minor comments (5)
  1. [Section IV-A] The MCC formula contains typographical errors: 'f p× f np' should be 'fp × fn', and the denominator is missing the square root; the sentence 'where is generally regarded' is also missing a subject.
  2. [Section IV-B] The text contains a typo: 'asa binary classification task' should be 'as a binary classification task'.
  3. [Throughout] The model name is inconsistently typeset as 'PedCA-FT', 'PEDCA-FT', and 'PEDCA-FT'; please standardize.
  4. [Section III-A and III-B] The notation '+ +' in Eqs. (8) and (9) is confusing; it is presumably meant to be string concatenation '++' or a named operator, and should be clarified.
  5. [Section II-B] The handling of admissions shorter than 24 hours is not described; please specify whether such admissions are excluded or how the feature window is truncated, since this interacts with the outcome-timeline concern in the major comments.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central predictive claim is derived empirically from a private cohort, and the self-citations present are background references, not load-bearing inputs.

full rationale

The paper's derivation chain is empirical rather than definitional: PedCA-FT maps EHR features x_{t<=24} to a risk estimate via learned transformer modules, and the reported metrics (balanced accuracy 63.08, F1 14.31, MCC 13.18, AUPRC 9.15, AUROC 73.99) come from 5-fold cross-validation in Section IV-A. The textualized view is a deterministic, lossy transformation of the same EHR features (Eqs. 7-9: min–max ranges for numerical time-series and value sets for categorical ones), so it is an input representation, not a label-derived quantity. The fusion model is trained with focal loss on the outcome labels; no fitted parameter is renamed as a prediction. The self-citations ([5], [6], [16], [22], [23]) are used for background statements such as 'data-driven approaches... leading to improved performance' or as method references, and none supplies the core result or forbids alternatives via a uniqueness theorem. The most serious concern in the paper is not circularity but a possible label-timeline ambiguity: Definition 2 predicts arrest 'during the rest of the admission stay' from data in the first 24 hours, while Section II-B does not explicitly state that patients who arrest within the first 24 hours are excluded; if they were included, the feature window could contain post-arrest measurements. That is a correctness/leakage risk, not a circular derivation, because the outcome is not defined in terms of the input features by the paper's equations. Overall, the derivation is self-contained against external benchmarks and no circular step reduces the central claim to its inputs.

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

The central claims rest on four domain assumptions and several unreported hyperparameters. None of the assumptions is machine-checked, and the outcome-definition assumption (no early-arrest contamination) is the most fragile.

free parameters (4)
  • PedCA-FT training hyperparameters (layers, hidden dimensions, learning rate, batch size, focal loss gamma) = Not reported
    Section III and IV-A describe architecture families but omit the concrete hyperparameter values; these choices determine the reported scores.
  • Missing-rate feature retention thresholds = 0.5 for vitals and labs, 0.7 for medications
    Section II-B selects 184 temporal risk factors using these thresholds; no sensitivity analysis is provided, and the thresholds shape the input space.
  • Prediction window length tau = 24 hours
    Definition 2 fixes the early-prediction horizon at 24 hours; all metrics are conditional on this choice.
  • Classification threshold used to compute F1, MCC, PPV, NPV, sensitivity, specificity = Not reported
    The paper reports threshold-dependent metrics but does not state the threshold or threshold-selection procedure for PedCA-FT; only baselines are described as using Top-K percentile based best threshold.
assumptions (4)
  • domain assumption The CHOA-CICU EHR extracts and CA labels are correctly curated.
    Section II-B describes the private database but provides no chart review, external validation, or label audit.
  • ad hoc to paper Patients who experience CA before the 24-hour feature cutoff are excluded or labeled without feature contamination.
    Definition 2 and Section II-B define the prediction window but never state the exclusion rule for early arrests; the model's scores depend on this.
  • domain assumption A pretrained BERT encoder transfers to short, template-based textualized EHR snippets.
    Section III-B assumes general text pretraining helps on the synthetic textual view, but no text-only ablation or domain adaptation is reported.
  • domain assumption External normal reference ranges used to filter abnormal labs (CALIPER-style) are appropriate for this pediatric CICU population.
    Section III-B filters out normal lab values using reference ranges; the ranges are cited but not verified for this cohort.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Early Risk Prediction of Pediatric Cardiac Arrest from Electronic Health Records via Multimodal Fused Transformer." pith.science (2026). https://pith.science/paper/OIXEKWBB

@misc{pith2026250207158,
  author       = {Pith},
  title        = {Pith review of: Early Risk Prediction of Pediatric Cardiac Arrest from Electronic Health Records via Multimodal Fused Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OIXEKWBB}},
  note         = {Machine review of arXiv:2502.07158}
}
read the original abstract

Early prediction of pediatric cardiac arrest (CA) is critical for timely intervention in high-risk intensive care settings. We introduce PedCA-FT, a novel transformer-based framework that fuses tabular view of EHR with the derived textual view of EHR to fully unleash the interactions of high-dimensional risk factors and their dynamics. By employing dedicated transformer modules for each modality view, PedCA-FT captures complex temporal and contextual patterns to produce robust CA risk estimates. Evaluated on a curated pediatric cohort from the CHOA-CICU database, our approach outperforms ten other artificial intelligence models across five key performance metrics and identifies clinically meaningful risk factors. These findings underscore the potential of multimodal fusion techniques to enhance early CA detection and improve patient care.

Figures

Figures reproduced from arXiv: 2502.07158 by the authors.

Figure 1
Figure 1. Model Architecture of PEDCA-FT. wide range of aggregation operation can be used for ϕ, and we opt to LAST(·) which keeps the last observed element from the whole time-series. The input of tabular-transformer Ttab is xt≤τ = (x (s) , x (t) agg) after applying Eq. (1). We adapt a tabular feature oriented Transformer [17], [18] variation as our tabular view. In a nutshell, the numerical and categorical inputs are first … view at source ↗
Figure 2
Figure 2. Bar plots with error bars illustrating PPV, NPV, Sensitivity, and Specificity. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Feature importance analysis using feature permutation. Error bar indicates p95 high and p95 low. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 29 canonical work pages

  1. [1]

    Risk factors and outcomes for recurrent paediatric in-hospital cardiac arrest: Retrospective multicenter cohort study,

    M. E. Frazier, S. R. Brown, A. O’Halloran, T. Raymond, R. Hanna, D. E. Niles, M. Kleinman, R. M. Sutton, J. Roberts, K. Tegtmeyer, et al., “Risk factors and outcomes for recurrent paediatric in-hospital cardiac arrest: Retrospective multicenter cohort study,” Resuscitation, vol. 169, pp. 60–66, 2021

  2. [2]

    Pic, a paediatric-specific intensive care database,

    X. Zeng, G. Yu, Y . Lu, L. Tan, X. Wu, S. Shi, H. Duan, Q. Shu, and H. Li, “Pic, a paediatric-specific intensive care database,” Scientific data, vol. 7, no. 1, p. 14, 2020

  3. [3]

    Pediatric in-hospital cardiac arrest and cardiopulmonary resuscitation in the united states: a review,

    R. W. Morgan, M. P. Kirschen, T. J. Kilbaugh, R. M. Sutton, and A. A. Topjian, “Pediatric in-hospital cardiac arrest and cardiopulmonary resuscitation in the united states: a review,” JAMA pediatrics , vol. 175, no. 3, pp. 293–302, 2021

  4. [4]

    B. D. Winters, M. M. Cvach, C. P. Bonafide, X. Hu, A. Konkani, M. F. O’Connor, J. M. Rothschild, N. M. Selby, M. M. Pelter, B. McLean, et al., “Society for critical care medicine alarm and alert fatigue task force. technological distractions (part 2): a summary of approaches to manage clinical alarms with intent to reduce alarm fatigue,” Crit Care Med, vo...

  5. [5]

    Prognostic accuracy of machine learning models for in-hospital mortality among children with phoenix sepsis admitted to the pediatric intensive care unit,

    R. Moore, D. Chanci, S. Brown, M. J. Ripple, N. R. Bishop, J. Grun- well, and R. Kamaleswaran, “Prognostic accuracy of machine learning models for in-hospital mortality among children with phoenix sepsis admitted to the pediatric intensive care unit,” Shock, vol. 63, no. 1, pp. 80–87, 2025

  6. [6]

    Using machine learning to predict cardiac arrest in the pediatric cardiac intensive care unit,

    S. Brown, J. Grunwell, Y . Wu, K. Dong, D. Bold, D. Liu, M. Fundora, X. Hu, and J. Lu, “Using machine learning to predict cardiac arrest in the pediatric cardiac intensive care unit,” Critical Care Medicine , vol. 53, no. 1, 2025

  7. [7]

    Ehrshot: An ehr benchmark for few-shot evaluation of foundation models,

    M. Wornow, R. Thapa, E. Steinberg, J. Fries, and N. Shah, “Ehrshot: An ehr benchmark for few-shot evaluation of foundation models,” Advances in Neural Information Processing Systems , vol. 36, pp. 67 125–67 137, 2023

  8. [8]

    Detecting and managing deterioration in children,

    A. Monaghan, “Detecting and managing deterioration in children,” Paediatric nursing, vol. 17, no. 1, p. 32, 2005

Show all 38 references
  1. [9]

    The inadequate oxygen delivery index and its correlation with venous saturation in the pediatric cardiac intensive care unit,

    R. S. Loomba, E. G. Villarreal, S. Flores, J. S. Farias, and A. Constas, “The inadequate oxygen delivery index and its correlation with venous saturation in the pediatric cardiac intensive care unit,” Pediatric cardiology, pp. 1–7, 2023

  2. [10]

    Xgboost: A scalable tree boosting system,

    T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , 2016, pp. 785–794

  3. [11]

    Scalable and accurate deep learning with electronic health records,

    A. Rajkomar, E. Oren, K. Chen, A. M. Dai, N. Hajaj, M. Hardt, P. J. Liu, X. Liu, J. Marcus, M. Sun, et al. , “Scalable and accurate deep learning with electronic health records,” NPJ digital medicine , vol. 1, no. 1, pp. 1–10, 2018

  4. [12]

    Moment: A family of open time-series foundation models,

    M. Goswami, K. Szafer, A. Choudhry, Y . Cai, S. Li, and A. Dubrawski, “Moment: A family of open time-series foundation models,” 2024. [Online]. Available: https://arxiv.org/abs/2402.03885

  5. [13]

    Lag-llama: Towards foundation models for time series forecasting,

    K. Rasul, A. Ashok, A. R. Williams, A. Khorasani, G. Adamopoulos, R. Bhagwatkar, M. Bilo ˇs, H. Ghonia, N. Hassen, A. Schneider, et al., “Lag-llama: Towards foundation models for time series forecasting,” in R0-FoMo: Robustness of Few-shot and Zero-shot Learning in Large Found...

  6. [14]

    Tiny time mixers (ttms): Fast pre-trained models for enhanced zero/few-shot forecasting of multivariate time series

    V . Ekambaram, A. Jati, P. Dayama, S. Mukherjee, N. H. Nguyen, W. M. Gifford, C. Reddy, and J. Kalagnanam, “Tiny time mixers (ttms): Fast pre-trained models for enhanced zero/few-shot forecasting of multivariate time series.” CoRR, 2024

  7. [15]

    Multimodal automl on structured tables with text fields,

    X. Shi, J. Mueller, N. Erickson, M. Li, and A. Smola, “Multimodal automl on structured tables with text fields,” in 8th ICML Workshop on Automated Machine Learning (AutoML) , 2021

  8. [16]

    MuG: A multimodal classification benchmark on game data with tabular, textual, and visual fields,

    J. Lu, Y . Qian, S. Zhao, Y . Xi, and C. Yang, “MuG: A multimodal classification benchmark on game data with tabular, textual, and visual fields,” in Findings of the Association for Computational Linguistics: EMNLP 2023, Dec. 2023, pp. 5332–5346

  9. [17]

    Tabtransformer: Tabular data modeling using contextual embeddings,

    X. Huang, A. Khetan, M. Cvitkovic, and Z. Karnin, “Tabtransformer: Tabular data modeling using contextual embeddings,” arXiv preprint arXiv:2012.06678, 2020

  10. [18]

    Revisiting deep learning models for tabular data,

    Y . Gorishniy, I. Rubachev, V . Khrulkov, and A. Babenko, “Revisiting deep learning models for tabular data,” Advances in Neural Informa- tion Processing Systems , vol. 34, pp. 18 932–18 943, 2021

  11. [19]

    A large language model for electronic health records,

    X. Yang, A. Chen, N. PourNejatian, H. C. Shin, K. E. Smith, C. Parisien, C. Compas, C. Martin, A. B. Costa, M. G. Flores, et al., “A large language model for electronic health records,” NPJ digital medicine, vol. 5, no. 1, p. 194, 2022

  12. [20]

    Dellir- ium: A large language model for delirium prediction in the icu using structured ehr,

    M. Contreras, S. Kapoor, J. Zhang, A. Davidson, Y . Ren, Z. Guan, T. Ozrazgat-Baslanti, S. Nerella, A. Bihorac, and P. Rashidi, “Dellir- ium: A large language model for delirium prediction in the icu using structured ehr,” arXiv preprint arXiv:2410.17363 , 2024

  13. [21]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems , vol. 30, 2017

  14. [22]

    Uncertainty-aware pre-trained foundation models for patient risk prediction via gaussian process,

    J. Lu, S. Zhao, W. Ma, H. Shao, X. Hu, Y . Xi, and C. Yang, “Uncertainty-aware pre-trained foundation models for patient risk prediction via gaussian process,” in Companion Proceedings of the ACM Web Conference 2024 , ser. WWW ’24, 2024, p. 1162–1165

  15. [23]

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

    Z. Chen, C. Ding, N. Modhe, J. Lu, C. Yang, and X. Hu, “Adapting a generative pretrained transformer achieves sota performance in assess- ing diverse physiological functions using only photoplethysmography signals: A gpt-ppg approach,” in AAAI 2024 Spring Symposium on Clinical...

  16. [24]

    Ehragent: Code empowers large language models for few-shot complex tabular reasoning on electronic health records,

    W. Shi, R. Xu, Y . Zhuang, Y . Yu, J. Zhang, H. Wu, Y . Zhu, J. Ho, C. Yang, and M. D. Wang, “Ehragent: Code empowers large language models for few-shot complex tabular reasoning on electronic health records,” in Proceedings of the 2024 Conference on Empirical Methods in Natur...

  17. [25]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. D. M.-W. C. Kenton and L. K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” in Proceedings of naacL-HLT, vol. 1. Minneapolis, Minnesota, 2019, p. 2

  18. [26]

    Clinical laboratory reference intervals in pediatrics: the caliper initiative,

    B. Jung and K. Adeli, “Clinical laboratory reference intervals in pediatrics: the caliper initiative,” Clinical biochemistry , vol. 42, no. 16-17, pp. 1589–1595, 2009

  19. [27]

    Focal loss for dense object detection,

    T.-Y . Ross and G. Doll ´ar, “Focal loss for dense object detection,” in proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2980–2988

  20. [28]

    Use of deep learning to develop continuous-risk models for adverse event prediction from electronic health records,

    N. Toma ˇsev, N. Harris, S. Baur, A. Mottram, X. Glorot, J. W. Rae, M. Zielinski, H. Askham, A. Saraiva, V . Magliulo, et al. , “Use of deep learning to develop continuous-risk models for adverse event prediction from electronic health records,” Nature Protocols, vol. 16, no. ...

  21. [29]

    Dynamic programming algorithm optimiza- tion for spoken word recognition,

    H. Sakoe and S. Chiba, “Dynamic programming algorithm optimiza- tion for spoken word recognition,” IEEE transactions on acoustics, speech, and signal processing , vol. 26, no. 1, pp. 43–49, 1978

  22. [30]

    Fast global alignment kernels,

    M. Cuturi, “Fast global alignment kernels,” in Proceedings of the 28th international conference on machine learning , 2011, pp. 929–936

  23. [31]

    Tslearn, a machine learning toolkit for time series data,

    R. Tavenard, J. Faouzi, G. Vandewiele, F. Divo, G. Androz, C. Holtz, M. Payne, R. Yurchak, M. Rußwurm, K. Kolar, and E. Woods, “Tslearn, a machine learning toolkit for time series data,” Journal of Machine Learning Research , vol. 21, no. 118, pp. 1–6, 2020. [Online]. Availabl...

  24. [32]

    Lightgbm: A highly efficient gradient boosting decision tree,

    G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T.-Y . Liu, “Lightgbm: A highly efficient gradient boosting decision tree,” in Advances in Neural Information Processing Systems , 2017, pp. 3146–3154

  25. [33]

    Tabnn: A uni- versal neural network solution for tabular data,

    G. Ke, J. Zhang, Z. Xu, J. Bian, and T.-Y . Liu, “Tabnn: A uni- versal neural network solution for tabular data,” arXiv preprint arXiv:1810.04805, 2018

  26. [34]

    Autogluon-tabular: Robust and accurate automl for structured data,

    N. Erickson, J. Mueller, A. Shirkov, H. Zhang, P. Larroy, M. Li, and A. Smola, “Autogluon-tabular: Robust and accurate automl for structured data,” arXiv preprint arXiv:2003.06505 , 2020

  27. [35]

    Time series classification from scratch with deep neural networks: A strong baseline,

    Z. Wang, W. Yan, and T. Oates, “Time series classification from scratch with deep neural networks: A strong baseline,” in 2017 International joint conference on neural networks (IJCNN). IEEE, 2017, pp. 1578– 1585

  28. [36]

    Pay attention to mlps,

    H. Liu, Z. Dai, D. So, and Q. V . Le, “Pay attention to mlps,” Advances in neural information processing systems , vol. 34, pp. 9204–9215, 2021

  29. [37]

    tsai - a state-of-the-art deep learning library for time series and sequential data,

    I. Oguiza, “tsai - a state-of-the-art deep learning library for time series and sequential data,” Github, 2023. [Online]. Available: https://github.com/timeseriesAI/tsai

  30. [38]

    Permutation importance: a corrected feature importance measure,

    A. Altmann, L. Tolos ¸i, O. Sander, and T. Lengauer, “Permutation importance: a corrected feature importance measure,” Bioinformatics, vol. 26, no. 10, pp. 1340–1347, 2010

Pith tools

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