Kidney Week

Abstract: SA-PO351

Proof-of-Concept Model for the Prediction of Dry Weight in Hemodialysis Patients

Session Information

• Hemodialysis and Frequent Dialysis: CV and Risk Prediction
  November 05, 2022 | Location: Exhibit Hall, Orange County Convention Center‚ West Building
  Abstract Time: 10:00 AM - 12:00 PM

Category: Dialysis

• 701 Dialysis: Hemodialysis and Frequent Dialysis

Authors

• Sandys, Vicki K., Royal College of Surgeons in Ireland, Dublin, Ireland

Vicki K. Sandys,

• Bhat, Lavleen, Royal College of Surgeons in Ireland, Dublin, Ireland

Lavleen Bhat,

• Sexton, Donal J., The University of Dublin Trinity College, Dublin, Ireland

Donal J. Sexton,

• O'Seaghdha, Conall M., Royal College of Surgeons in Ireland, Dublin, Ireland

Conall M. O'Seaghdha,

Group or Team Name

• Haemodialysis Outcomes & Patient Empowerment group

Background

An automated, accurate and periodic assessment of dry weight using hemodialysis data would be a clinically useful, low cost, and rapidly scalable method of assessing fluid status.

Methods

A post-hoc analysis of a 3 week observational study in 20 patients (HOPE-02, ClinicalTrials.gov: NCT04623281). 143 features were created using clinical and calculated parameters over 1 or 2 weeks prior to the targeted session including variability measures, differences, confidence intervals and interdialytic calculations. Averaged data was combined with single-session data. Dry weight was defined as the mean of BCM normohydration weights. Data was split 70:30 into training and testing sets.The algorithm was trained to predict overhydration (OH) index pre and post-dialysis as determined by the Fresenius Body Composition Monitor (BCM). Dry weight was derived by subtracting OH status from pre or post-dialysis weight. Model performance was evaluated using adjusted R². The error metrics used were mean absolute error (MAE) and root mean squared error (RMSE). Accuracy was calculated using categories of fluid status defined as overhydration > 1.1 L, normohydration 1.1 - 1.1 L and underhydration < 1.1 L.

Results

20 subjects involving 44 haemodialysis sessions were used. A linear regression (LR) model combining single-session data with moving averages of data from the preceding 1 or 2 weeks was created. The final model had 9 features (Figure 1). The LR model predicting post-dialysis OH status outperformed a pre-dialysis OH model. Adjusted R² for the model was 0.719. Training RMSE for post-dialysis OH= 1.12 kg, testing RMSE= 1.21 kg. The training RMSE for dry weight= 1.21 kg, the testing RMSE= 1.13 kg. The model had a classification accuracy of 81.8%.

Conclusion

This model represents an initial step in the creation of an automated assessment of dry weight. Work has begun on external validation and further development of this algorithm in a large dialysis dataset. A final model could potentially be embedded in a clinical decision support system for dry weight management.

Funding

• Commercial Support – Enteprise Ireland Disruptive Technologies Innovation Fund grant DTIF 2019_86