Abstract: PO0576
Development of a Machine Learning Approach to Management of CKD-MBD Therapy
Session Information
- Bone and Mineral Metabolism: Clinical
November 04, 2021 | Location: On-Demand, Virtual Only
Abstract Time: 10:00 AM - 12:00 PM
Category: Bone and Mineral Metabolism
- 402 Bone and Mineral Metabolism: Clinical
Authors
- Gaweda, Adam E., University of Louisville School of Medicine, Louisville, Kentucky, United States
- Brier, Michael E., University of Louisville School of Medicine, Louisville, Kentucky, United States
- Lederer, Eleanor D., VA North Texas Health Care System, Dallas, Texas, United States
Background
We developed a Quantitative Systems Pharmacology (QSP) model of CKD-MBD that predicts changes in mineral metabolism. We incorporate the CKD-MBD model into an Machine Learning (ML)-based simulation to optimize the dosing of three drugs used in CKD-MBD to test the hypothesis that Reinforcement Learning (RL) approach would improve therapeutic goals.
Methods
We performed a simulated 24 month study in a virtual cohort of 80 Stage 5d CKD patients using the QSP model of CKD-MBD treated by a simulated physician (AI-Agent 0) or RL (AI-Agent 1). Agent 0 was a Deep Neural Network trained on a set of 128,061 instances. Agent 1 was developed using RL rewarding concentrations within the target range for Ca, P, PTH and avoiding Ca < 7.0 and > 10.2 mg/dL. Results of the simulation were compared using regression analysis of the dependent variable (Ca, P, PTH, calcitriol (CTL), lnFGF23, bCa(bone efflux), and vCa(vascular influx) over time with the factors RL (Agent1 vs Agent 0), P binder adherence, and equilibrium vs. steady-state. Doses of agents used to treat were compared at 24 months.
Results
Results of the statitical analysis are shown in the Table. Agent 1 using RL resulted in a greater rate of change in the dependent variables in all cases and resulted in lower model predicted concentrations of P, PTH, and FGF23 and higher concentrations of Ca, and CTL. The time effect on FGF23 was not significant. Ca flux from the bone and into the tissue was also decreased in Agent 1. Drug utilization was also different between methods tested at 24 months. Agent 1: 734 mg/day less P binder (p=0.015), 0.71 ug/day more calcitriol (p<0.001) and 3.84 mg/day less cinecalcet (p<0.001).
Conclusion
Through simulation we have shown that a machine learning approach using reinforcement learning is superior to an expert system mimicking physician dosing practices. Concentrations of Ca, P, and PTH came into equlibrium faster and at more optimal levels while predicting decreased unwanted Ca movement.
| Coefficient | p Value | |||||
| Time | RL | Adherance | Time | RL | Adherance | |
| P | -0.053 | -0.167 | -0.603 | <0.001 | <0.001 | <0.001 |
| Ca | 0.053 | 0.237 | 0.155 | <0.001 | <0.001 | <0.001 |
| PTH | -24.2 | -53.7 | -81.6 | <0.001 | <0.001 | <0.001 |
| CTL | 1.53 | 19.6 | -7.76 | <0.001 | <0.001 | <0.001 |
| lnFGF23 | 0.004 | -0.329 | -1.34 | 0.435 | <0.001 | <0.001 |
| vCa | -0.006 | -0.017 | -0.118 | <0.001 | <0.001 | <0.001 |
| bCa | -0.017 | -0.117 | -0.08 | <0.001 | <0.001 | <0.001 |
Funding
- Veterans Affairs Support