Abstract: PO0100
Federated Learning for AKI Prediction in COVID-19 Patients
Session Information
- COVID-19: Health Systems and More
November 04, 2021 | Location: On-Demand, Virtual Only
Abstract Time: 10:00 AM - 12:00 PM
Category: Coronavirus (COVID-19)
- 000 Coronavirus (COVID-19)
Authors
- Jaladanki, Suraj K., Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Vaid, Akhil, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Paranjpe, Ishan, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Saha, Aparna, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Singh, Karandeep, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Chan, Lili, Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Glicksberg, Benjamin S., Icahn School of Medicine at Mount Sinai, New York, New York, United States
- Nadkarni, Girish N., Icahn School of Medicine at Mount Sinai, New York, New York, United States
Background
Predictive models are trained on single-center data and are non-generalizable, and multi-center data pooling raises privacy concerns. Federated learning (FL) trains models by updating parameters from a central aggregator without sharing raw data. We used FL to predict acute kidney injury (AKI) in COVID-19 patients within 3 (AKI3) and 7 (AKI7) days of admission as a use case.
Methods
We selected 4035 COVID-19 patients admitted to 5 hospitals in New York City, after excluding patients with kidney failure, to train logistic regression and logistic regression with L1 regularization (LASSO) models through 3 approaches: local data, pooled data from all sites, and a FL method.
Results
Federated models outperformed local models as measured by area under the receiver operating characteristic curve (Figure 1, Table 1). SHAP plots indicate differences in feature importance between LASSO models in AKI3 prediction (Figure 2).
Conclusion
FL has utility for developing accurate predictive models without compromising patient data.
Table 1.
| Classifier | Model | Mount Sinai Brooklyn (AKI3: 36.6%, AKI7: 44.2%, N = 658) | Mount Sinai Hospital (AKI3: 29.6%; AKI7: 35.2%, N = 1445) | Mount Sinai Morningside (AKI3: 34.0%, AKI7: 40.4%, N = 810) | Mount Sinai Queens (AKI3: 33.0%, AKI7: 39.8%, N = 648) | Mount Sinai West (AKI3: 23.4%, AKI7: 26.6%, N = 474) | Cross Site Average | |
| AKI within 3 Days (AKI3) | Logistic Regression (LR) | Local | 0.797 (0.792 - 0.803) | 0.794 (0.791 - 0.798) | 0.782 (0.777 - 0.787) | 0.766 (0.760 - 0.771) | 0.798 (0.790 - 0.805) | 0.787 (0.785 - 0.790) |
| Pooled | 0.831 (0.826 - 0.835) | 0.840 (0.837 - 0.843) | 0.818 (0.814 - 0.822) | 0.850 (0.846 - 0.854) | 0.874 (0.869 - 0.878) | 0.843 (0.840 - 0.845) | ||
| Federated | 0.777 (0.772 - 0.783) | 0.802 (0.798 - 0.805) | 0.794 (0.790 - 0.799) | 0.822 (0.818 - 0.827) | 0.852 (0.847 - 0.857) | 0.810 (0.807 - 0.812) | ||
| Logistic Regression with L1 Regularization (LASSO) | Local | 0.800 (0.794 - 0.805) | 0.789 (0.785 - 0.793) | 0.783 (0.778 - 0.788) | 0.761 (0.755 - 0.767) | 0.801 (0.794 - 0.808) | 0.787 (0.784 - 0.790) | |
| Pooled | 0.835 (0.830 - 0.839) | 0.840 (0.837 - 0.844) | 0.825 (0.821 - 0.829) | 0.855 (0.851 - 0.859) | 0.872 (0.867 - 0.877) | 0.846 (0.843 - 0.848) | ||
| Federated | 0.778 (0.773 - 0.783) | 0.802 (0.799 - 0.806) | 0.795 (0.791 - 0.800) | 0.822 (0.818 - 0.827) | 0.852 (0.847 - 0.857) | 0.810 (0.807 - 0.813) | ||
| AKI within 7 Days (AKI7) | Logistic Regression (LR) | Local | 0.790 (0.785 - 0.794) | 0.782 (0.779 - 0.785) | 0.771 (0.767 - 0.774) | 0.743 (0.739 - 0.748) | 0.781 (0.776 - 0.785) | 0.773 (0.771 - 0.775) |
| Pooled | 0.831 (0.827 - 0.834) | 0.830 (0.828 - 0.832) | 0.810 (0.807 - 0.813) | 0.821 (0.817 - 0.824) | 0.857 (0.854 - 0.861) | 0.830 (0.828 - 0.831) | ||
| Federated | 0.783 (0.779 - 0.787) | 0.796 (0.793 - 0.798) | 0.791 (0.787 - 0.794) | 0.792 (0.788 - 0.796) | 0.843 (0.840 - 0.847) | 0.801 (0.799 - 0.803) | ||
| Logistic Regression with L1 Regularization (LASSO) | Local | 0.792 (0.787 - 0.796) | 0.774 (0.771 - 0.777) | 0.771 (0.767 - 0.774) | 0.739 (0.735 - 0.744) | 0.786 (0.781 - 0.791) | 0.772 (0.770 - 0.775) | |
| Pooled | 0.833 (0.829 - 0.836) | 0.827 (0.825 - 0.829) | 0.814 (0.811 - 0.817) | 0.822 (0.818 - 0.825) | 0.858 (0.854 - 0.862) | 0.831 (0.829 - 0.832) | ||
| Federated | 0.783 (0.779 - 0.787) | 0.796 (0.793 - 0.798) | 0.791 (0.788 - 0.794) | 0.792 (0.788 - 0.796) | 0.844 (0.841 - 0.847) | 0.801 (0.799 - 0.803) |
Average model performance across hospitals by area under the receiver-operating characteristic curve.
SHAP plots of LASSO local and federated models in predicting AKI within 3 days of admission at Mount Sinai Hospital.
Funding
- NIDDK Support