Real-time prediction of intradialytic hypotension using machine learning and cloud computing infrastructure

Abstract

Background: In maintenance hemodialysis patients, intradialytic hypotension (IDH) is a frequent complication that has been associated with poor clinical outcomes. Prediction of IDH may facilitate timely interventions and eventually reduce IDH rates.

Methods: We developed a machine learning model to predict IDH in in-center hemodialysis patients 15-75 min in advance. IDH was defined as systolic blood pressure (SBP) <90 mmHg. Demographic, clinical, treatment-related and laboratory data were retrieved from electronic health records and merged with intradialytic machine data that were sent in real-time to the cloud. For model development, dialysis sessions were randomly split into training (80%) and testing (20%) sets. The area under the receiver operating characteristic curve (AUROC) was used as a measure of the model's predictive performance.

Results: We utilized data from 693 patients who contributed 42 656 hemodialysis sessions and 355 693 intradialytic SBP measurements. IDH occurred in 16.2% of hemodialysis treatments. Our model predicted IDH 15-75 min in advance with an AUROC of 0.89. Top IDH predictors were the most recent intradialytic SBP and IDH rate, as well as mean nadir SBP of the previous 10 dialysis sessions.

Conclusions: Real-time prediction of IDH during an ongoing hemodialysis session is feasible and has a clinically actionable predictive performance. If and to what degree this predictive information facilitates the timely deployment of preventive interventions and translates into lower IDH rates and improved patient outcomes warrants prospective studies.

Keywords: end-stage kidney disease; intradialytic hypotension; machine learning; real-time prediction.

Graphical Abstract

Figure 1:

Evaluation of model performance. The performance analysis comprised two aspects, one per SBP measurement (usually multiple SBP measurements per HD session) and one that considered the entire HD session. Evaluation of model performance was based on the definitions of TP, FP, TN and FN.

Figure 2:

(A) AUROC. The true positive rate is shown on the y-axis, equal to sensitivity. The false positive rate on the x-axis is calculated as 1 – specificity. The 95% CIs of the AUROC are 0.881–0.892. (B) Calibration plot with equal bin width. Probability calibration plot shows the predicted probability against observed events. (C) Calibration plot with equal number of samples per bin (200 bins in total). Here, instead of equal bin-width, we set the bin width based on the number of samples to account for the data distribution (most probabilities are under 0.5).

Figure 3:

Top 10 predictors for IDH in the ML model. (A) Mean absolute SHAP values. (B) The SHAP summary plot shows the degree of each measurement's positive or negative effect on the prediction (x-axis). Warmer colors represent higher observed values for that measurement; cooler colors indicate lower values. For example, the higher (warmer color) the “most recent SBP,” the more the negative impact it has on the model (less chance of IDH).

Figure 4:

Two examples demonstrating a patient with (A; left panel) and without (B; right panel) IDH. (A) An example of SBP (top row) and IDH probability (bottom row) in a patient who experienced IDH. The model predicted the probability of IDH throughout HD with each SBP measurement (approximately every 30 min during a regular HD treatment); the model was trained to predict IDH 15–75 min before the event. An IDH probability ≥0.09 was set as IDH alert threshold (dashed horizonal lines in the lower panels). Data after the IDH event were not considered. (B) An example of SBP and predicted IDH probability in a patient who did not experience IDH; the IDH probability remained below the IDH alert threshold throughout the treatment.