Abstract: FR-PO0016
Comparison of Causal Machine Learning (ML) Models on Predicting Survival Benefit with Dialysis in Patients with ESKD: A Retrospective Cohort Study
Session Information
- Artificial Intelligence: Clinical Prediction, Decision Support, and Digital Kidney Care
October 23, 2026 | Location: Exhibit Hall A, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Artificial Intelligence, Digital Health, and Data Science
- 300 Artificial Intelligence, Digital Health, and Data Science
Authors
- Leung, Ka Chun, Tuen Mun Hospital, Hong Kong, Hong Kong
- Yap, Yat Hin Desmond, The University of Hong Kong, Hong Kong, Hong Kong
Background
The choice to start dialysis at the bedside in advanced CKD is highly individualized, but average-effect trial evidence (e.g., IDEAL) cannot tell patient-specific treatment benefit. Causal ML methods offer a potential solution for quantifying this heterogeneity from observational data, but clinicians lack clear evidence on which method to trust.
Methods
Eight causal-ML models were trained on a Hong Kong CKD cohort (n = 20184) and validated on a external cohort (n = 6040). Primary outcome: IPCW-weighted mortality at 1 year. Model evaluation included discrimination, calibration, permutation null test and transportability across cohorts.
Results
The permutation null test rejected no informative HTE on every model (p ≤ 0.025) except the IPCW_DML_Pseudo model. All models preserved C-for-benefit 0.53 – 0.55 (risk-proportional baseline 0.513), proving causal ML edge over a clinician's heuristic. Treating the top 30 % by predicted CATE rank recovered 75 % (95 % CI 59 – 98 %) of the treat-all mortality benefit on the external test set, while sparing 70 % of patients from early dialysis.
Conclusion
Causal-ML models produce population-level treatment effects and detect treatment-effect heterogeneity (p < 0.025), while preserve their patient-level ranking across different cohorts. Calibration and transportability are the dimensions that distinguish outstanding models at the bedside.
Per-method summary on external test set 1-year mortality
| Method | C-for-benefit | Δ vs no-HTE null | Calibration slope | Global O:E | Q1–Q5 spread | Recovery at K = 30 % | Transport ΔC, spatial → external |
| Stacked ensemble | 0.554 (0.528-0.564) | +0.036 (+0.015-+0.064) | 0.77 (0.52-0.95) | 0.56 | 0.184 (0.117-0.258) | 75 (59-98) | -0.013 (-0.050-+0.011) |
| S-learner | 0.543 (0.519-0.561) | +0.028 (+0.009-+0.059) | 0.62 (0.41-0.82) | 0.57 | 0.218 (0.144-0.296) | 73 (59-93) | -0.025 (-0.050-+0.022) |
| X-learner | 0.541 (0.525-0.559) | +0.026 (+0.007-+0.059) | 0.72 (0.50-0.92) | 0.55 | 0.189 (0.116-0.256) | 73 (59-89) | -0.033 (-0.065-+0.005) |
| IPTW-S-learner | 0.545 (0.527-0.560) | +0.033 (+0.009-+0.060) | 1.03 (0.70-1.40) | 0.71 | 0.166 (0.106-0.230) | 69 (54-89) | -0.027 (-0.062-+0.002) |
| Survival-DML-Cox | 0.540 (0.523-0.558) | +0.026 (+0.007-+0.058) | 0.52 (0.31-0.73) | 0.64 | 0.190 (0.108-0.282) | 65 (51-85) | -0.019 (-0.057-+0.021) |
| T-learner | 0.547 (0.524-0.557) | +0.038 (+0.008-+0.056) | 0.47 (0.28-0.64) | 0.57 | 0.158 (0.102-0.227) | 64 (51-79) | -0.016 (-0.064-+0.009) |
| Kennedy-DR-learner | 0.533 (0.514-0.548) | +0.017 (-0.002-+0.050) | 0.46 (0.26-0.67) | 0.55 | 0.138 (0.048-0.203) | 54 (37-68) | -0.015 (-0.056-+0.018) |
| IPCW-DML pseudo-outcome | 0.506 (0.494-0.531) | -0.005 (-0.020-+0.027) | 0.23 (0.05-0.40) | 0.54 | 0.095 (0.035-0.180) | 47 (28-69) | -0.049 (-0.086--0.017) |
Discrimination per method on external validation
Recovery curves: fraction of treat-all mortality benefit recovered when treating only the top K % of patients by predicted CATE rank