gms | German Medical Science

38. Jahrestagung der Deutschsprachigen Arbeitsgemeinschaft für Verbrennungsbehandlung (DAV 2020)

15.01. - 18.01.2020, Zell am See, Österreich

Artikel

Artikel empfehlen

Using AI to predict the length of hospitalization in pediatric burn patients

Meeting Abstract

Suche in Medline nach

  • J. Elrod - Department of Pediatric Surgery, Burn Unit, Plastic and Reconstructive Surgery, Altonaer Children’s Hospital, Hamburg, Germany
  • C. Mohr - Department of Pediatric Surgery, Burn Unit, Plastic and Reconstructive Surgery, Altonaer Children’s Hospital, Hamburg, Germany
  • R. Wolff - Neoglia LTD, London, United Kingdom
  • K. Reinshagen - Department of Pediatric Surgery, Burn Unit, Plastic and Reconstructive Surgery, Altonaer Children’s Hospital, Hamburg, Germany; Department of Pediatric Surgery, University Medical Center Hamburg-Eppendorf, Hamburg, Germany
  • Committee of the German Burn Registry
  • Ingo Königs - Department of Pediatric Surgery, Burn Unit, Plastic and Reconstructive Surgery, Altonaer Children’s Hospital, Hamburg, Germany; Department of Pediatric Surgery, University Medical Center Hamburg-Eppendorf, Hamburg, Germany

Deutschsprachige Arbeitsgemeinschaft für Verbrennungsbehandlung. 38. Jahrestagung der Deutschsprachigen Arbeitsgemeinschaft für Verbrennungsbehandlung (DAV 2020). Zell am See, Österreich, 15.-18.01.2020. Düsseldorf: German Medical Science GMS Publishing House; 2020. Doc11.04

doi: 10.3205/20dav080, urn:nbn:de:0183-20dav0804

Veröffentlicht: 13. Januar 2020

© 2020 Elrod et al.
Dieser Artikel ist ein Open-Access-Artikel und steht unter den Lizenzbedingungen der Creative Commons Attribution 4.0 License (Namensnennung). Lizenz-Angaben siehe http://creativecommons.org/licenses/by/4.0/.

Gliederung

Text

Introduction: Prediction of length of stay (LOS) is of great importance in burn patients.

This study aims at investigating the prediction accuracy of complex models and a linear regression-based approach. The heuristic expecting 1 day of stay per % total body surface area (TBSA) commonly applied as a rule of thumb in burn care is used as the performance benchmark.

Methods: The study is based on pediatric burn patient data sets provided by the German Society for Burn Treatment (DGV) with N=8542. Mean absolute and standard error (MAE resp. MSE) are calculated for each prediction model (rule of thumb, linear regression and random forest) and the level of significance is investigated. To test the statistical difference between the selected estimators, 20-fold cross validation is performed. In addition, the relationship between TBSA and the residual error is analyzed and factors contributing to a prolonged stay are determined.

Results: The present analysis indicates the prediction accuracy (MAE, MSE) of the two methods of interest to be statistically significantly superior to the rule of thumb (p < 0.01). Furthermore, the residual error rises as TBSA increases for all three methods. Factors associated with a prolonged LOS are % TBSA by degree, inhalation trauma and scald.

Conclusion: More complex models lead to a moderate increase in prediction accuracy in this cohort. Our work indicates that improving the predictive power of such machine learning systems will require recording more details about each burn incident. Generally speaking, the use of artificial intelligence for the purpose of data analysis could lead to more evidence in medicine.