41250 Machine Learning to Identify Predictors of Iatrogenic Injury Using Empirical Bayes Estimates: A Cohort Study of Pressure Injury Prevention

Padula, WV; Armstrong, DG; Davidson, PM

41250 Machine Learning to Identify Predictors of Iatrogenic Injury Using Empirical Bayes Estimates: A Cohort Study of Pressure Injury Prevention

Padula, WV Armstrong, DG Davidson, PM

Permalink

Publisher:: Cambridge University Press (CUP)
Publication Type:: Journal Article
Citation:: Journal of Clinical and Translational Science, 2021, 5, (Suppl 1), pp. 49-49
Issue Date:: 2021-03-30

Open Access

Copyright Clearance Process

Recently Added
In Progress
Open Access

This item is open access.

Adobe PDF

Download Published versionAdobe PDF (52.09 kB)

View on publisher's site

View statistics

Full metadata record

Field	Value	Language
dc.contributor.author	Padula, WV
dc.contributor.author	Armstrong, DG
dc.contributor.author	Davidson, PM
dc.date.accessioned	2022-03-24T05:38:53Z
dc.date.available	2022-03-24T05:38:53Z
dc.date.issued	2021-03-30
dc.identifier.citation	Journal of Clinical and Translational Science, 2021, 5, (Suppl 1), pp. 49-49
dc.identifier.issn	2059-8661
dc.identifier.issn	2059-8661
dc.identifier.uri	http://hdl.handle.net/10453/155520
dc.description.abstract	<jats:p>ABSTRACT IMPACT: A machine learning approach using electronic health records can combine descriptive, population-level factors of pressure injury outcomes. OBJECTIVES/GOALS: Pressure injuries cause 60,000 deaths and cost $26 billion annually in the US, but prevention is laborious. We used clinical data to develop a machine learning algorithm for predicting pressure injury risk and prescribe the timing of intervention to help clinicians balance competing priorities. METHODS/STUDY POPULATION: We obtained 94,745 electronic health records with 7,000 predictors to calibrate a predictive algorithm of pressure injury risk. Machine learning was used to mine features predicting changes in pressure injury risk; random forests outperformed neural networks, boosting and bagging in feature selection. These features were fit to multilevel ordered logistic regression to create an algorithm that generated empirical Bayes estimates informing a decision-rule for follow-up based on individual risk trajectories over time. We used cross-validation to verify predictive validity, and constrained optimization to select a best-fit algorithm that reduced the time required to trigger patient follow-up. RESULTS/ANTICIPATED RESULTS: The algorithm significantly improved prediction of pressure injury risk (p<0.001) with an area under the ROC curve of 0.60 compared to the Braden Scale, a traditional clinician instrument of pressure injury risk. At a specificity of 0.50, the model achieved a sensitivity of 0.63 within 2.5 patient-days. Machine learning identified categorical increases in risk when patients were prescribed vasopressors (OR=16.4, p<0.001), beta-blockers (OR=4.8, p<0.001), erythropoietin stimulating agents (OR=3.0, p<0.001), or were ordered a urinalysis screen (OR=9.1, p<0.001), lipid panel (OR=5.7, p<0.001) or pre-albumin panel (OR=2.0, p<0.001). DISCUSSION/SIGNIFICANCE OF FINDINGS: This algorithm could help hospitals conserve resources within a critical period of patient vulnerability for pressure injury not reimbursed by Medicare. Savings generated by this approach could justify investment in machine learning to develop electronic warning systems for many iatrogenic injuries.</jats:p>
dc.language	en
dc.publisher	Cambridge University Press (CUP)
dc.relation.ispartof	Journal of Clinical and Translational Science
dc.relation.isbasedon	10.1017/cts.2021.530
dc.rights	info:eu-repo/semantics/openAccess
dc.title	41250 Machine Learning to Identify Predictors of Iatrogenic Injury Using Empirical Bayes Estimates: A Cohort Study of Pressure Injury Prevention
dc.type	Journal Article
utslib.citation.volume	5
pubs.organisational-group	/University of Technology Sydney
pubs.organisational-group	/University of Technology Sydney/Faculty of Health
utslib.copyright.status	open_access	*
dc.date.updated	2022-03-24T05:38:52Z
pubs.issue	Suppl 1
pubs.publication-status	Published
pubs.volume	5
utslib.citation.issue	Suppl 1

Abstract:

ABSTRACT IMPACT: A machine learning approach using electronic health records can combine descriptive, population-level factors of pressure injury outcomes. OBJECTIVES/GOALS: Pressure injuries cause 60,000 deaths and cost $26 billion annually in the US, but prevention is laborious. We used clinical data to develop a machine learning algorithm for predicting pressure injury risk and prescribe the timing of intervention to help clinicians balance competing priorities. METHODS/STUDY POPULATION: We obtained 94,745 electronic health records with 7,000 predictors to calibrate a predictive algorithm of pressure injury risk. Machine learning was used to mine features predicting changes in pressure injury risk; random forests outperformed neural networks, boosting and bagging in feature selection. These features were fit to multilevel ordered logistic regression to create an algorithm that generated empirical Bayes estimates informing a decision-rule for follow-up based on individual risk trajectories over time. We used cross-validation to verify predictive validity, and constrained optimization to select a best-fit algorithm that reduced the time required to trigger patient follow-up. RESULTS/ANTICIPATED RESULTS: The algorithm significantly improved prediction of pressure injury risk (p<0.001) with an area under the ROC curve of 0.60 compared to the Braden Scale, a traditional clinician instrument of pressure injury risk. At a specificity of 0.50, the model achieved a sensitivity of 0.63 within 2.5 patient-days. Machine learning identified categorical increases in risk when patients were prescribed vasopressors (OR=16.4, p<0.001), beta-blockers (OR=4.8, p<0.001), erythropoietin stimulating agents (OR=3.0, p<0.001), or were ordered a urinalysis screen (OR=9.1, p<0.001), lipid panel (OR=5.7, p<0.001) or pre-albumin panel (OR=2.0, p<0.001). DISCUSSION/SIGNIFICANCE OF FINDINGS: This algorithm could help hospitals conserve resources within a critical period of patient vulnerability for pressure injury not reimbursed by Medicare. Savings generated by this approach could justify investment in machine learning to develop electronic warning systems for many iatrogenic injuries.

Please use this identifier to cite or link to this item:

http://hdl.handle.net/10453/155520