Abstract: TH-OR19
Multicenter Development and Validation of a Multimodal Deep Learning Model to Predict Severe AKI
Session Information
- Augmented Intelligence and Digital Health Advances
October 24, 2024 | Location: Room 2, Convention Center
Abstract Time: 04:30 PM - 04:40 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Authors
- Koyner, Jay L., University of Chicago Division of the Biological Sciences, Chicago, Illinois, United States
- Martin, Jennifer, University of Wisconsin-Madison, Madison, Wisconsin, United States
- Carey, Kyle, University of Chicago Division of the Biological Sciences, Chicago, Illinois, United States
- Caskey, John, University of Wisconsin-Madison, Madison, Wisconsin, United States
- Dligach, Dmtriy, University of Wisconsin-Madison, Madison, Wisconsin, United States
- Afshar, Majid, University of Wisconsin-Madison, Madison, Wisconsin, United States
- Churpek, Matthew M., University of Wisconsin-Madison, Madison, Wisconsin, United States
Background
Electronic health record(EHR) -based risk algorithms for acute kidney injury(AKI) have been reported but these have focused on structured data(e.g. vital signs & laboratory values). We aimed to develop & validate a deep learning time series model to predict AKI by combining structured data & unstructured data from clinical notes
Methods
Patients(≥18 years) admitted to the University of Wisconsin (2009-20) & the University of Chicago Medicine (2016-22) were included. Patients were excluded if they had no documented serum creatinine (SCr), an admission SCr≥3.0mg/dL, developed Stage 2 AKI before reaching the wards or intensive care unit (ICU), or if they required dialysis(RRT) within the first 48 hrs. AKI was defined by the KDIGO SCr definition. Fifty-eight structured features were used in the model, and the raw text from unstructured clinical notes were mapped to standardized Concept Unique Identifiers (CUIs) to create unstructured data features. An intermediate fusion deep learning recurrent neural network architecture was used to predict Stage ≥2 AKI within the next 48 hrs. The model was developed in the first 80% of the data by date & temporally validated in the next 20%.
Results
There were 339,998 subjects in the derivation cohort & 84,581 in the validation cohort with 12,748(3%) across both cohorts, developing Stage 2 AKI. Those with Stage 2 AKI were more likely to be older, male, have higher baseline SCr, and more likely to be in the ICU. CUIs related to intubation, ventilation, and airway management were more common in those with AKI. AUCs overall and across important subgroups were above 0.85 for most outcomes.
Conclusion
We developed and validated a novel deep-learning risk assessment tool using structured & unstructured data from the EHR in real time to predict the development of Stage 2 AKI across the entire hospital. A real-time version of this model has now been implemented at both health systems.
AUC for AKI Outcomes in the Next 48 hours
| All Values AUC(95%CI) | At Least Stage 2 AKI | Receipt of RRT |
| All Encounters | 0.87 (0.87-0.87) | 0.90 (0.90-0.90) |
| Ward Patients | 0.86 (0.86-0.86) | 0.88 (0.87-0.88) |
| ICU Patients | 0.83 (0.83-0.83) | 0.84 (0.84-0.84 |
| Baseline Creatinine < 1.0 mg/dL | 0.86 (0.86-0.87) | 0.89 (0.89-0.90) |
| Baseline Creatinine 1.0-1.99 mg/dL | 0.88 (0.88-0.88) | 0.90 (0.90-0.90) |
| Baseline Creatinine >2.0 mg/dL | 0.88 (0.88-0.88) | 0.87 (0.87-0.87) |
Funding
- NIDDK Support