**Authors and Affiliations**
Authors: Kuroush Nezafati, MD#1, Sreekanth Cheruku, MD#2, Donghan M. Yang, PhD1, Tingyi Wanyan, PhD1, Qinbo Zhou, PhD1, Javier A. Neyra, MD, MSCS3, Xiaowei Zhan, PhD*1, Yang Xie, PhD*1, Amanda Fox, MD, MPH*2
(#: equal contributions; *: co-corresponding)
1. Quantitative Biomedical Research Center, Peter O’Donell School of Public Health, University of Texas Southwestern Medical Center 5323 Harry Hines Blvd, Dallas, TX, USA 75390;
2. Department of Anesthesiology UT Southwestern Medical Center, 5323 Harry Hines Blvd, Dallas, TX, USA 75390
3. Division of Nephrology The University of Alabama at Birmingham, 1900 University Boulevard, THT 647, Birmingham, AL, USA 35233

**Background**
Cardiac Surgery Associated Acute Kidney Injury (CSA-AKI) is a serious complication of cardiac surgery, affecting up to 30% of patients. Given the significant associated morbidity and mortality with CSA-AKI, accurate AKI prediction for patients undergoing cardiac surgery is highly valued.

**Methods**
574 patients with Coronary Artery Bypass Graft (CABG) or valve surgery or both surgeries were enrolled. All patients had blood samples (Nephrocheck® and BNP) drawn pre and postoperatively for biomarker analysis. Patients had basic pre-operative labs. Vital data (e.g. measurements from Swan-Ganz catheter, arterial blood pressure, pulse, etc.) measured from start of surgery through hour 12 in the ICU. A deep learning based long short-term memory (LSTM)model and random forest (RF) model were developed and evaluated using a cohort of 459 patients. The outcome of prediction was CSA-AKI as determined by change in creatinine within 7-day postoperative window. Results reported on a randomly selected, independent testing set of remaining 115 patients within the cohort. Finally, we created a baseline random forest model without time-series data to see this data’s significance.

**Results**
In predicting CSA-AKI, the LSTM model using both time-series vital data and biomarkers for the AKI prediction had an accuracy of 0.84 and AUC of 0.71. This surpassed the performances from the RF model (accuracy was 0.77 and AUC 0.66). Our baseline random forest model only achieved suboptimal performance (accuracy = 0.61; AUC = 0.47).

**Conclusion**
A LSTM model incorporating both time-series vital data and AKI biomarkers showed greater performance at predicting CSA-AKI over a more traditional machine learning method in RF. We also showed the importance of time-series vital data for the prediction of AKI.



**Authors and Affiliations**
**Authors:** Kuroush Nezafati, MD^#1^, Sreekanth Cheruku, MD^#2^, Donghan M. Yang, PhD^1^, Tingyi Wanyan, PhD^1^, Qinbo Zhou, PhD^1^, Javier A. Neyra, MD, MSCS^3^, Xiaowei Zhan, PhD^*1^, Yang Xie, PhD^*1^, Amanda Fox, MD, MPH^*2^
(#: equal contributions; *: co-corresponding)
1. Quantitative Biomedical Research Center, Peter O’Donell School of Public Health, University of Texas Southwestern Medical Center 5323 Harry Hines Blvd, Dallas, TX, USA 75390;
2. Department of Anesthesiology UT Southwestern Medical Center, 5323 Harry Hines Blvd, Dallas, TX, USA 75390
3. Division of Nephrology The University of Alabama at Birmingham, 1900 University Boulevard, THT 647, Birmingham, AL, USA 35233

**Background**
Cardiac Surgery Associated Acute Kidney Injury (CSA-AKI) is a serious complication of cardiac surgery, affecting up to 30% of patients. Given the significant associated morbidity and mortality with CSA-AKI, accurate AKI prediction for patients undergoing cardiac surgery is highly valued.

**Methods**
574 patients with Coronary Artery Bypass Graft (CABG) or valve surgery or both surgeries were enrolled. All patients had blood samples (Nephrocheck® and BNP) drawn pre and postoperatively for biomarker analysis. Patients had basic pre-operative labs. Vital data (e.g. measurements from Swan-Ganz catheter, arterial blood pressure, pulse, etc.) measured from start of surgery through hour 12 in the ICU. A deep learning based long short-term memory (LSTM)model and random forest (RF) model were developed and evaluated using a cohort of 459 patients. The outcome of prediction was CSA-AKI as determined by change in creatinine within 7-day postoperative window. Results reported on a randomly selected, independent testing set of remaining 115 patients within the cohort. Finally, we created a baseline random forest model without time-series data to see this data’s significance.

**Results**
In predicting CSA-AKI, the LSTM model using both time-series vital data and biomarkers for the AKI prediction had an accuracy of 0.84 and AUC of 0.71. This surpassed the performances from the RF model (accuracy was 0.77 and AUC 0.66). Our baseline random forest model only achieved suboptimal performance (accuracy = 0.61; AUC = 0.47).

**Conclusion**
A LSTM model incorporating both time-series vital data and AKI biomarkers showed greater performance at predicting CSA-AKI over a more traditional machine learning method in RF. We also showed the importance of time-series vital data for the prediction of AKI.



Predicting Acute Kidney Injury in Cardiac Surgical Patients with Time Series Analysis Using Deep Learning Model