Abstract: TH-PO034
Using Artificial Intelligence (AI) to Predict Mortality in AKI Patients: A Systematic Review and Meta-Analysis
Session Information
- AI, Digital Health, Data Science - I
November 02, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Augmented Intelligence, Digital Health, and Data Science
- 300 Augmented Intelligence, Digital Health, and Data Science
Authors
- Raina, Rupesh, Cleveland Clinic Akron General, Akron, Ohio, United States
- Shah, Raghav, Northeast Ohio Medical University, Rootstown, Ohio, United States
- Sethi, Sidharth Kumar, Medanta The Medicity Medanta Institute of Kidney and Urology, Gurugram, Haryana, India
- Koyner, Jay L., The University of Chicago Medicine, Chicago, Illinois, United States
- Neyra, Javier A., Division of Nephrology, University of Alabama at Birmingham, Birmingham, Alabama, United States
Background
Acute kidney injury (AKI) is associated with increased morbidity and mortality. With the recent advent of artificial intelligence (AI), novel models for mortality prediction in AKI patients have been developed using machine learning (ML). We reviewed the performance of different ML models for evidence generation to support their applicability and implementation in the clinical setting.
Methods
A literature search was conducted through Pubmed, Embase, and Web of Science Databases. Performance metrics of the ML models to predict hospital mortality in adult AKI patients were extracted. The between-study heterogeneity was assessed using the I2 test and random [for I2 ≥ 50%] and fixed effects model were used. The AUROC of two models were compared using DeLong’s test (p-value ≤ 0.05 considered significant). R software version 3.1.0 was used in the analysis.
Results
A total of 8 studies [8 derivation and 6 validation cohorts] with 37,032 adult AKI patients were included. The hospital mortality was 18% in the derivation and 15.8% in the validation cohorts. The pooled AUC (95% CI) was observed to be highest for logistic regression [0.86 (0.80 - 0.93)] and lowest for proposed clinical [0.77 (0.72 - 0.81)] models used as reference. Despite substantial variability, the pooled AUC (95% CI) of logistic regression did not differ significantly from other models except proposed clinical model [Delong’s test p=0.022].
Conclusion
Our results show that logistic regression is equally effective as other ML models in predicting in-hospital mortality among AKI patients, with substantial variability across models. Studies evaluating the features influencing mortality and their impact on different models are needed.
Meta-analysis of AUC in assessing in-hospital mortality among AKI patients
| AI/ML Model | No. of Mortality / Sample size | No of cohorts; no of studies | Pooled AUC (95% CI) | I2 (95% CI); p values |
| Logistic regression | 1,225/5,690 | 3 D and 3 V cohort; 3 studies | 0.862 (0.795 - 0.928) | 96.17% (93.77% - 97.64%); p<0.0001 |
| Broad learning system models* | 278/540 | 2 D and 2 V cohort; 1 study | 0.852 (0.820 - 0.883) | 44.88% (0.00% - 81.61%); p=0.1421 |
| Elastic Net final model fitted* | 544/870 | 1 D and 1 V cohort; 1 study | 0.852 (0.813 - 0.891) | 48.05% (8.04% - 75.25%); p=0.1653 |
| Extreme gradient boost | 2,525/5,642 | 3 D and 2 V cohort; 3 studies | 0.816 (0.752 - 0.880) | 95.09% (91.20% - 97.26%); p<0.0001 |
| Random Forest | 5,613/27,467 | 7 D and 5 V cohort; 6 studies | 0.816 (0.782 - 0.851) | 90.14% (84.72% - 93.64%); p<0.0001 |
| Support vector machine | 4,571/23,816 | 3 D and 1 V cohort; 3 studies | 0.804 (0.760 - 0.849) | 93.89% (87.53% - 97.01%); p<0.0001 |
| ANN / MLP | 4,571/23,816 | 3 D and 1 V cohort; 3 studies | 0.793 (0.752 - 0.833) | 87.22% (69.38% - 94.66%); p<0.0001 |
| Proposed clinical model | 1,831/9,587 | 1 D and 1 V cohort; 1 study | 0.765 (0.716 - 0.814) | 98.96% (97.92% - 99.48%); p<0.0001 |
*Fixed effect models, for other random effect models D:Derivation, V:Validation, ANN:Artificial Neural Network, MLP:Multi-layer perceptron