Kidney Week

Abstract: FR-PO414

Prediction of Gastrointestinal Bleeding Hospitalization Risk in Hemodialysis: Machine Learning vs. Logistic Regression

Session Information

• Hemodialysis Epidemiology and Outcomes
  October 25, 2024 | Location: Exhibit Hall, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Dialysis

• 801 Dialysis: Hemodialysis and Frequent Dialysis

Authors

• Lama, Suman Kumar, Fresenius Medical Care, Waltham, Massachusetts, United States

Suman Kumar Lama, MSc

• Larkin, John W., Fresenius Medical Care, Waltham, Massachusetts, United States

John W. Larkin, PhD

• Chaudhuri, Sheetal, Fresenius Medical Care, Waltham, Massachusetts, United States

Sheetal Chaudhuri, PhD

• Jiao, Yue, Fresenius Medical Care, Waltham, Massachusetts, United States

Yue Jiao, PhD

• Winter, Anke, Fresenius Medical Care Germany, Bad Homburg, Germany

Anke Winter, MD

• Stauss-Grabo, Manuela, Fresenius Medical Care Germany, Bad Homburg, Germany

Manuela Stauss-Grabo, PhD

• Usvyat, Len A., Fresenius Medical Care, Waltham, Massachusetts, United States

Len A. Usvyat, PhD

• Hymes, Jeffrey L., Fresenius Medical Care, Waltham, Massachusetts, United States

Jeffrey L. Hymes, MD

• Maddux, Franklin W., Fresenius Medical Care AG, Bad Homburg, Hessen, Germany

Franklin W. Maddux, MD

• Wheeler, David C., University College London, London, United Kingdom

David C. Wheeler, MD

• Stenvinkel, Peter, Karolinska Universitetsjukhuset Huddinge Njurmedicinska kliniken, Stockholm, Sweden

Peter Stenvinkel, MD, PhD, FASN

• Floege, Jürgen, Rheinisch-Westfalische Technische Hochschule Aachen, Aachen, Nordrhein-Westfalen, Germany

Jürgen Floege, MD

Group or Team Name

• On behalf of the INSPIRE Core Group.

Background

Gastrointestinal bleeding (GIB) is the most common bleeding event in dialysis. INSPIRE group aimed to see if machine learning (ML) models (XGBoost and logistic regression) can identify GIB related hospitalization risk in hemodialysis (HD) patients.

Methods

We used data from adult dialysis patients treated for ≥30 days at a kidney care network (Jan2018-Mar2021). GIB hospitalization was defined based on ICD diagnosis codes recorded as primary, secondary, or tertiary discharge reason. Two distinct models were created using XGBoost and logistic regression on same dataset. Performance was evaluated using area under receiver operating curve (AUROC), accuracy, sensitivity and specificity. Data was randomly divided into 60% training, 20% test and 20% validation datasets. We used unseen patients and data in test data to evaluate the performance of the model.

Results

Incidence rate of 180-day GIB hospitalization was 1.12% in our test data. AUROC was 0.74 for XGBoost and 0.68 for logistic regression. XGBoost achieved a specificity of 0.68 and a sensitivity of 0.65, while logistic regression demonstrated a specificity of 0.57 and a sensitivity of 0.69. Both models identified age, distribution in anemia/iron indices, recent all-cause hospitalization and bone mineral metabolism markers as strong predictor.

Conclusion

ML modeling has potential to detect of GIB hospitalization risk in hemodialysis patients. XGBoost model outperforms logistic regression, yet both performed well. The link between bone mineral metabolism markers and GI bleeding was not anticipated and merits further investigation. Additional validation may be required to ensure model reliability before implementation in a clinical setting.

Figure 1 AUROC of GI Bleed model with XGBoost and Logistic Regression model