Kidney Week

Abstract: TH-OR21

Hb Co-Pilot: Machine Learning Algorithm for Real-Time Hemoglobin Estimation during Hemodialysis and a Multicenter International Validation Trial

Session Information

• Augmented Intelligence and Digital Health Advances
  October 24, 2024 | Location: Room 2, Convention Center
  Abstract Time: 04:50 PM - 05:00 PM

Category: Augmented Intelligence, Digital Health, and Data Science

• 300 Augmented Intelligence, Digital Health, and Data Science

Authors

• Kuo, Chin-Chi, Big Data Center, China Medical University Hospital, Taichung, Taiwan

Chin-Chi Kuo, MD, PhD, MPH

• Chen, Sheng-Hsuan, Big Data Center, China Medical University Hospital, Taichung, Taiwan

Sheng-Hsuan Chen, RD

• Chiang, Hsiu-Yin, Big Data Center, China Medical University Hospital, Taichung, Taiwan

Hsiu-Yin Chiang, MS, PhD

• Sun, Chuan-Hu, Big Data Center, China Medical University Hospital, Taichung, Taiwan

Chuan-Hu Sun, MS

• Hsieh, Mei-Chuan, Big Data Center, China Medical University Hospital, Taichung, Taiwan

Mei-Chuan Hsieh

• Chou, Che-yi, Asia University Hospital, Taichung, Taiwan

Che-yi Chou, MD, PhD

• Holt, Stephen Geoffrey, SEHA Kidney Care, Abu Dhabi, Abu Dhabi, United Arab Emirates

Stephen Geoffrey Holt, MBBS, PhD

Background

Anemia is linked to premature death among patients receiving hemodialysis (HD). We developed and validated a timely, non-invasive, and simple smartphone APP (Hb Co-Pilot) that captures images of HD dialysis tube and predicts hemoglobin (Hb) level using a machine-learning model.

Methods

Training & Testing Set We enrolled adult HD patients with arteriovenous fistula (AVF) or graft at China Medical University Healthcare System (CMUHS). HD tubing photos on Fresenius 4008(s) machine were taken with a color correction matrix (CCM) on the day of blood testing. A total of 5,453 images were taken by 13 smartphones from 5 HD centers.
Feature Extraction & Modeling Images were pre-processed to ensure similar lighting conditions. We cropped the areas of HD tube and CCM and extracted the 192-dimension vector of image features from both areas by encoding the color information in histograms using a bin size of 64 on each RGB channel. An XGBoost model was used to predict Hb >10 g/dL using image features, age, gender, and last Hb level, with a testing accuracy of 0.93.
Validation Trial CMUHS, Asia University Hospital (AUH) in Taiwan, and SEHA Kidney Care (SKC) in Abu Dhabi participated in our validation trial, with a total of 504 images collected. The targeted area under the receiver operating curve (AUROC) was set to 0.80.

Results

In the validation study, the median age was 62 years, 67% were male, 40% from SKC, 85% had AVF, 26% had Hb ≤10 g/dL, and 40% were taken from Fresenius 5008. The mean of difference between predicted and true Hb was 0.83 g/dL. The AUROC and accuracy was both 0.81, and the sensitivity and positive predictive value both reached 0.87. The F1 and Kappa was 0.87 and 0.50, respectively.

Conclusion

Hb Co-Pilot had robust external validity in Taiwanese and Emirati across different dialysis machines. Model improvement will focus on capturing extreme Hb levels and evaluating its clinical effectiveness.

Funding

• Government Support – Non-U.S.