Kidney Week

Abstract: PUB250

Exploring Therapeutic Effects of Continuous Kidney Replacement Therapy in Patients with Severe Acidosis Using Deep Learning-Based Causal Inference

Session Information

• Publication Only

Category: Fluid, Electrolytes, and Acid-Base Disorders

• 1102 Fluid, Electrolyte, and Acid-Base Disorders: Clinical

Authors

• Kang, Min Woo, Seoul National University College of Medicine, Jongno-gu, Seoul, Korea (the Republic of)

Min Woo Kang

• Park, Sehoon, Seoul National University College of Medicine, Jongno-gu, Seoul, Korea (the Republic of)

Sehoon Park, MD, PhD

• Kim, Yong Chul, Seoul National University College of Medicine, Jongno-gu, Seoul, Korea (the Republic of)

Yong Chul Kim, MD, PhD

• Oh, Kook-Hwan, Seoul National University College of Medicine, Jongno-gu, Seoul, Korea (the Republic of)

Kook-Hwan Oh, MD, PhD

• Kim, Dong Ki, Seoul National University College of Medicine, Jongno-gu, Seoul, Korea (the Republic of)

Dong Ki Kim, MD, PhD

Background

Continuous kidney replacement therapy (CKRT) is an essential treatment for uncontrolled severe metabolic acidosis. However, CKRT can increase workload and lead to complications, thus necessitating its selective application to patients who stand to benefit significantly. This study aims to investigate the therapeutic effect of CKRT in patients with severe acidosis by utilizing a deep learning-based causal inference model to assess its potential impact on in-hospital mortality.

Methods

The Medical Information Mart for Intensive Care III (MIMIC-III) dataset was utilized, and patients with data available within the first 48 hours after intensive care unit (ICU) admission were selected. Patients experiencing severe acidosis with a pH <7.2 within the initial 48 hours were selected. Treatment was defined as the application of CKRT within 48 hours of ICU admission, and the outcome was defined as in-hospital mortality. The dataset was randomly divided into an 85:15 ratio for train and test data. The Generative Adversarial Nets for Inference of Individualized Treatment Effects (GANITE) model was applied on training the model, and its performance was evaluated in the test data.

Results

In the train data, the model demonstrated an accuracy and Area Under Receiver Operating Characteristic Curve (AUROC) of 0.883 and 0.887 (0.880–0.893), respectively, while in the test data, it showed 0.841 and 0.824 (0.804–0.843). The model was well calibrated. The probability change of average in-hospital mortality with CKRT treatment for all severe acidosis patients was +15% and +14% in the train and test data, respectively. However, in the patient who underwent CKRT, the application of CKRT resulted in an average reduction of in-hospital mortality probability by 13% in both the train and test data.

Conclusion

Developing a model that strategically represents the therapeutic effects of CKRT for individual patients could be expected to aid decision-making in the future application of CKRT treatment.

Examples