pith. sign in

arxiv: 1907.11195 · v1 · pith:YTVYDUBPnew · submitted 2019-07-25 · 📊 stat.ML · cs.LG· stat.AP

Deep Learning Models to Predict Pediatric Asthma Emergency Department Visits

Xiao Wang , Zhijie Wang , Yolande M. Pengetnze , Barry S. Lachman , Vikas Chowdhry This is my paper

Pith reviewed 2026-05-24 15:56 UTC · model grok-4.3

classification 📊 stat.ML cs.LGstat.AP
keywords pediatric asthmaemergency department visitsartificial neural networksLasso logistic regressionMedicaid claims dataprediction modelsAUC performancedeep learning applications
0
0 comments X

The pith

An artificial neural network slightly outperforms Lasso logistic regression at predicting which children with asthma will visit the emergency department within three months.

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

The paper trains an artificial neural network on Medicaid claims records to forecast asthma-related emergency department visits in the next three months and compares it with a penalized logistic regression model already deployed since 2015. The network reaches an AUC of 0.845 while the regression model reaches 0.842, with the difference attributed to the network's capacity to model nonlinear relationships among the predictors. A sympathetic reader would care because pediatric asthma affects millions of children and better forecasts could allow earlier interventions that reduce emergency visits, medication adjustments, and family stress. The work demonstrates that routine billing data already contain enough signal for deep learning to deliver a measurable, if modest, gain over linear methods in this setting.

Core claim

The central claim is that an artificial neural network model, trained on Medicaid claims data, predicts asthma-related emergency department visits within three months with an area under the curve of 0.845, slightly higher than the 0.842 obtained by the Lasso logistic regression model that has been in production since 2015, and that the improvement arises from the network's ability to capture nonlinear patterns in the data.

What carries the argument

Artificial neural network applied to Medicaid claims features for three-month-ahead binary prediction of asthma emergency department visits, benchmarked against Lasso logistic regression.

If this is right

  • If the performance edge holds, neural network models can replace or supplement the existing Lasso model for risk stratification in pediatric asthma programs.
  • Higher prediction accuracy would support more precise targeting of education, trigger avoidance, and medication reviews for children most likely to need emergency care.
  • The modest gain illustrates that nonlinear models can extract additional value from the same claims features that linear models already use.
  • Deployment of such models could improve resource allocation within Medicaid managed-care organizations by identifying high-risk patients earlier.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The small difference in AUC raises the question of whether larger or richer datasets would widen the gap or whether simpler models remain adequate for this task.
  • Extending the same approach to non-Medicaid populations or to other chronic childhood conditions could test how far claims-based deep learning generalizes.
  • Pairing the model with real-time environmental or pharmacy refill data might produce larger accuracy improvements than architecture changes alone.
  • If the model were run prospectively in clinical workflows, the practical value would depend on whether the risk scores actually change clinician or family behavior.

Load-bearing premise

The Medicaid claims data contain all relevant predictors, are free of systematic missingness or coding bias, and the three-month prediction window plus train-test split produce an unbiased estimate of future performance.

What would settle it

A prospective evaluation on a fresh cohort of pediatric Medicaid patients in which the artificial neural network's AUC is no higher than the Lasso model's or in which the model's risk scores do not correlate with actual future emergency department visits.

Figures

Figures reproduced from arXiv: 1907.11195 by Barry S. Lachman, Vikas Chowdhry, Xiao Wang, Yolande M. Pengetnze, Zhijie Wang.

Figure 1
Figure 1. Figure 1: Asthma cohort selection process In order to compare with our baseline model in production, the data for this study was extracted from PCHP claims data between July 2012 and June 2014, which was the same time range as the one we used to train the Lasso logistic regression model. We first filtered for children aged between 6 months and 18 years old at prediction time and applied CSTE “probable” asthma criter… view at source ↗
Figure 2
Figure 2. Figure 2: ANN connection in our model desired binary outcome. We define the loss function and backpropagate the error to hidden layers to update the weights. We iterate this process until the predefined convergence rate is achieved. 3.3 Training and Test Strategies Data was randomly divided into training and test sets in 0.7/0.3 proportions. We applied various resampling methods to obtain different training samples … view at source ↗
Figure 3
Figure 3. Figure 3: ROC curve and PR curve of ANN metrics and clinical assessment of the likely capacity of a potential intervention program, we designed the patients in the top most 10 percent of risk scores as “High", the 10th to 20th percentile range as “Medium” and the rest as “Low” risk in the Lasso logistic regression model. We kept the same thresholds for ANN model. The predicted adverse events were only from “High” ri… view at source ↗
read the original abstract

Pediatric asthma is the most prevalent chronic childhood illness, afflicting about 6.2 million children in the United States. However, asthma could be better managed by identifying and avoiding triggers, educating about medications and proper disease management strategies. This research utilizes deep learning methodologies to predict asthma-related emergency department (ED) visit within 3 months using Medicaid claims data. We compare prediction results against traditional statistical classification model - penalized Lasso logistic regression, which we trained and have deployed since 2015. The results have indicated that deep learning model Artificial Neural Networks (ANN) slightly outperforms (with AUC = 0.845) the Lasso logistic regression (with AUC = 0.842). The reason may come from the nonlinear nature of ANN.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript claims that an artificial neural network (ANN) model for predicting pediatric asthma emergency department visits within a 3-month horizon using Medicaid claims data achieves a slightly higher AUC (0.845) than a penalized Lasso logistic regression model (0.842) that has been deployed since 2015, attributing the difference to the nonlinear modeling capacity of the ANN.

Significance. If the performance difference were shown to be statistically reliable and the evaluation free of leakage or label noise, the work would provide modest evidence that deep learning can extract additional signal from claims data beyond linear penalized models. The practical significance remains limited by the 0.003 AUC gap and the absence of any demonstration that the improvement translates to better clinical decision-making or reduced ED utilization.

major comments (2)
  1. [Abstract] Abstract: The central claim that ANN 'slightly outperforms' Lasso rests on an AUC difference of 0.003 with no accompanying standard error, bootstrap interval, DeLong test, or any other statistical comparison; in claims-data settings the typical AUC variance after temporal splits is 0.01–0.03, so the reported gap is compatible with noise and does not support the outperformance conclusion.
  2. [Methods] Methods/Results: No information is given on total sample size, number of predictors after preprocessing, train/test split sizes, cross-validation procedure, or handling of censoring and missing claims codes; without these quantities the reported AUCs cannot be interpreted or compared.
minor comments (1)
  1. [Abstract] The abstract states that the Lasso model 'we trained and have deployed since 2015' but provides no follow-up metrics on its real-world calibration or drift; adding such information would strengthen the baseline comparison.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment below and have revised the manuscript to incorporate additional statistical comparisons and methodological details.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim that ANN 'slightly outperforms' Lasso rests on an AUC difference of 0.003 with no accompanying standard error, bootstrap interval, DeLong test, or any other statistical comparison; in claims-data settings the typical AUC variance after temporal splits is 0.01–0.03, so the reported gap is compatible with noise and does not support the outperformance conclusion.

    Authors: We agree that the AUC difference of 0.003 is small and that the absence of a statistical comparison prevents any firm conclusion of outperformance. In the revised manuscript we have added bootstrap confidence intervals for both AUC estimates and included the result of a DeLong test for the paired ROC curves. The abstract and discussion have been updated to report the AUC values without claiming superiority of the ANN model. revision: yes

  2. Referee: [Methods] Methods/Results: No information is given on total sample size, number of predictors after preprocessing, train/test split sizes, cross-validation procedure, or handling of censoring and missing claims codes; without these quantities the reported AUCs cannot be interpreted or compared.

    Authors: We acknowledge that these essential details were omitted from the original submission. The revised Methods section now reports the total sample size, the number of predictors retained after preprocessing, the sizes of the training and test sets under the temporal split, the cross-validation procedure employed for model tuning, and the approach taken to censoring and missing claims codes. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical model comparison on external claims data

full rationale

The paper reports AUCs from training and evaluating ANN and Lasso logistic regression on Medicaid claims data for a 3-month prediction task. No derivation chain exists that reduces a claimed result to its own fitted parameters by construction, nor any self-citation load-bearing step, uniqueness theorem, or ansatz smuggling. The comparison is standard supervised learning with train/test split; reported numbers are direct outputs of model fitting rather than renamed inputs. This is the most common non-circular case for applied ML papers.

Axiom & Free-Parameter Ledger

3 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard supervised-learning assumptions plus several modeling choices whose justification is not visible in the abstract.

free parameters (3)
  • ANN architecture and hyperparameters
    Number of layers, hidden units, learning rate, and regularization strength are chosen to maximize performance on the training data.
  • Lasso penalty strength
    Regularization parameter selected during model training on the same claims data.
  • 3-month prediction horizon
    Time window chosen by the authors; not derived from first principles.
axioms (2)
  • domain assumption Claims records contain all clinically relevant predictors and are missing at random conditional on observed features.
    Invoked implicitly when treating the feature matrix as sufficient for prediction.
  • domain assumption The held-out test set is exchangeable with future patients.
    Required for the reported AUC to be interpreted as prospective performance.

pith-pipeline@v0.9.0 · 5669 in / 1473 out tokens · 20595 ms · 2026-05-24T15:56:30.228415+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages · 2 internal anchors

  1. [1]

    Ahmadizar, S

    F. Ahmadizar, S. J. Vijverberg, H. G. Arets, A. de Boer, J. E. Lang, M. Kattan, C. N. Palmer, S. Mukhopadhyay, S. Turner, and A. H. Maitland-van der Zee. Childhood obesity in relation to poor asthma control and exacerbation: a meta-analysis. European Respiratory Journal, 48(4):1063–1073, 2016

    work page 2016

  2. [2]

    A. L. Andrews, A. N. Simpson, W. T. Basco Jr, R. J. Teufel, et al. Asthma medication ratio predicts emergency department visits and hospitalizations in children with asthma. Medicare & medicaid research review, 3(4), 2013

    work page 2013

  3. [3]

    M. A. Badgeley, J. R. Zech, L. Oakden-Rayner, B. S. Glicksberg, M. Liu, W. Gale, M. V . McConnell, B. Percha, T. M. Snyder, and J. T. Dudley. Deep learning predicts hip fracture using confounding patient and healthcare variables. npj Digital Medicine, 2(1):31, 2019

    work page 2019

  4. [4]

    W. E. Berger, A. P. Legorreta, M. S. Blaiss, E. C. Schneider, A. T. Luskin, D. A. Stempel, S. Suissa, D. C. Goodman, S. W. Stoloff, J. A. Chapman, et al. The utility of the health plan employer data and information set (hedis) asthma measure to predict asthma-related outcomes. Annals of Allergy, Asthma & Immunology , 93(6):538–545, 2004

    work page 2004

  5. [5]

    M. S. Broder, B. Gutierrez, E. Chang, D. Meddis, and M. Schatz. Ratio of controller to total asthma medications: determinants of the measure. The American journal of managed care, 16(3):170–178, 2010

    work page 2010

  6. [6]

    CDC - asthma - data and surveillance - asthma surveillance data, 2018

    CDC.gov. CDC - asthma - data and surveillance - asthma surveillance data, 2018

    work page 2018

  7. [7]

    CDC - asthma, 2019

    CDC.gov. CDC - asthma, 2019

    work page 2019

  8. [8]

    G. Daniel. Principles of artificial neural networks, volume 7. World Scientific, 2013

    work page 2013

  9. [9]

    L. T. Das, E. L. Abramson, A. E. Stone, J. E. Kondrich, L. M. Kern, and Z. M. Grinspan. Predicting frequent emergency department visits among children with asthma using ehr data. Pediatric pulmonology, 52(7):880–890, 2017

    work page 2017

  10. [10]

    Dreiseitl and L

    S. Dreiseitl and L. Ohno-Machado. Logistic regression and artificial neural network classification models: a methodology review. Journal of biomedical informatics, 35(5-6):352–359, 2002

    work page 2002

  11. [11]

    Esteva, A

    A. Esteva, A. Robicquet, B. Ramsundar, V . Kuleshov, M. DePristo, K. Chou, C. Cui, G. Corrado, S. Thrun, and J. Dean. A guide to deep learning in healthcare. Nature medicine, 25(1):24, 2019

    work page 2019

  12. [12]

    Forno, A

    E. Forno, A. Fuhlbrigge, M. E. Soto-Quirós, L. Avila, B. A. Raby, J. Brehm, J. M. Sylvia, S. T. Weiss, and J. C. Celedón. Risk factors and predictive clinical scores for asthma exacerbations in childhood. Chest, 138(5):1156– 1165, 2010

    work page 2010

  13. [13]

    E. W. Gelfand, G. L. Colice, L. Fromer, W. B. Bunn III, and T. J. Davies. Use of the health plan employer data and information set for measuring and improving the quality of asthma care. Annals of Allergy, Asthma & Immunology, 97(3):298–305, 2006

    work page 2006

  14. [14]

    Goodfellow, Y

    I. Goodfellow, Y . Bengio, and A. Courville.Deep Learning. MIT Press, 2016. http://www.deeplearningbook. org

    work page 2016

  15. [15]

    D. K. Greineder, K. C. Loane, and P. Parks. A randomized controlled trial of a pediatric asthma outreach program. Journal of Allergy and Clinical Immunology, 103(3):436–440, 1999

    work page 1999

  16. [16]

    W. J and B. A. The benefit of using both claims data and electronic medical record data in health care analysis. Technical report, Optum Insight, 2012

    work page 2012

  17. [17]

    Karnick, H

    P. Karnick, H. Margellos-Anast, G. Seals, S. Whitman, G. Aljadeff, and D. Johnson. The pediatric asthma intervention: a comprehensive cost-effective approach to asthma management in a disadvantaged inner-city community. Journal of Asthma, 44(1):39–44, 2007

    work page 2007

  18. [18]

    S. M. Lauritsen, M. E. Kalør, E. L. Kongsgaard, K. M. Lauritsen, M. J. Jørgensen, J. Lange, and B. Thiesson. Early detection of sepsis utilizing deep learning on electronic health record event sequences. arXiv preprint arXiv:1906.02956, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1906

  19. [19]

    LeCun, Y

    Y . LeCun, Y . Bengio, and G. Hinton. Deep learning.nature, 521(7553):436, 2015

    work page 2015

  20. [20]

    Miotto, F

    R. Miotto, F. Wang, S. Wang, X. Jiang, and J. T. Dudley. Deep learning for healthcare: review, opportunities and challenges. Briefings in bioinformatics, 19(6):1236–1246, 2017

    work page 2017

  21. [21]

    D. M. Mosen, E. Macy, M. Schatz, G. Mendoza, T. B. Stibolt, J. McGaw, J. Goldstein, and J. Bellows. How well do the hedis asthma inclusion criteria identify persistent asthma. Am J Manag Care, 11(10):650–4, 2005

    work page 2005

  22. [22]

    Nurmagambetov, O

    T. Nurmagambetov, O. Khavjou, L. Murphy, and D. Orenstein. State-level medical and absenteeism cost of asthma in the united states. Journal of Asthma, 54(4):357–370, 2017. 6 Deep Learning Models to Predict Pediatric Asthma Emergency Department Visits

    work page 2017

  23. [23]

    Nurmagambetov, R

    T. Nurmagambetov, R. Kuwahara, and P. Garbe. The economic burden of asthma in the united states, 2008–2013. Annals of the American Thoracic Society, 15(3):348–356, 2018

    work page 2008

  24. [24]

    Processing of Electronic Health Records using Deep Learning: A review

    V . Osmani, L. Li, M. Danieletto, B. Glicksberg, J. Dudley, and O. Mayora. Processing of electronic health records using deep learning: A review. arXiv preprint arXiv:1804.01758, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  25. [25]

    C. M. Pacheco, C. E. Ciaccio, N. Nazir, C. M. Daley, A. DiDonna, W. S. Choi, C. S. Barnes, and L. J. Rosenwasser. Homes of low-income minority families with asthmatic children have increased condition issues. In Allergy and asthma proceedings, volume 35, page 467. OceanSide Publications, 2014

    work page 2014

  26. [26]

    T. Pham, T. Tran, D. Phung, and S. Venkatesh. Deepcare: A deep dynamic memory model for predictive medicine. In Pacific-Asia Conference on Knowledge Discovery and Data Mining, pages 30–41. Springer, 2016

    work page 2016

  27. [27]

    B. K. Reddy and D. Delen. Predicting hospital readmission for lupus patients: An rnn-lstm-based deep-learning methodology. Computers in biology and medicine, 101:199–209, 2018

    work page 2018

  28. [28]

    Schatz, E

    M. Schatz, E. F. Cook, A. Joshua, and D. Petitti. Risk factors for asthma hospitalizations in a managed care organization: development of a clinical prediction rule. The American journal of managed care, 9(8):538–547, 2003

    work page 2003

  29. [29]

    Schatz, R

    M. Schatz, R. S. Zeiger, W. M. V ollmer, D. Mosen, G. Mendoza, A. J. Apter, T. B. Stibolt, A. Leong, M. S. Johnson, and E. F. Cook. The controller-to-total asthma medication ratio is associated with patient-centered as well as utilization outcomes. Chest, 130(1):43–50, 2006

    work page 2006

  30. [30]

    R. H. Stanford, M. B. Shah, A. O. D’Souza, and M. Schatz. Predicting asthma outcomes in commercially insured and medicaid populations? The American journal of managed care, 19(1):60–67, 2013

    work page 2013

  31. [31]

    Tolomeo, C

    C. Tolomeo, C. Savrin, M. Heinzer, and A. Bazzy-Asaad. Predictors of asthma-related pediatric emergency department visits and hospitalizations. Journal of Asthma, 46(8):829–834, 2009

    work page 2009

  32. [32]

    D. B. Wakefield and M. M. Cloutier. Modifications to hedis and cste algorithms improve case recognition of pediatric asthma. Pediatric pulmonology, 41(10):962–971, 2006

    work page 2006

  33. [33]

    M. Xu, K. G. Tantisira, A. Wu, A. A. Litonjua, J.-h. Chu, B. E. Himes, A. Damask, and S. T. Weiss. Genome wide association study to predict severe asthma exacerbations in children using random forests classifiers.BMC medical genetics, 12(1):90, 2011

    work page 2011

  34. [34]

    H. S. Zahran, C. M. Bailey, S. A. Damon, P. L. Garbe, and P. N. Breysse. Vital signs: asthma in children—united states, 2001–2016. Morbidity and Mortality Weekly Report, 67(5):149, 2018. 7

    work page 2001