Abstract: TH-PO203
Using Machine Learning to Help Predict Elevated Serum Phosphate Levels in Patients with ESRD
Session Information
- Bone and Mineral Metabolism: Clinical - I
October 25, 2018 | Location: Exhibit Hall, San Diego Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Bone and Mineral Metabolism
- 402 Bone and Mineral Metabolism: Clinical
Authors
- Long, Andrew, Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Blanchard, Tommy C., Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Willetts, Joanna, Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Chaudhuri, Sheetal, Fresenius Medical Care North America, Waltham, Massachusetts, United States
- O'Connell, Michael Ryan, Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Belmonte, Kathleen, Fresenius Kidney Care, Waltham, Massachusetts, United States
- Lee, Marissa A., Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Usvyat, Len A., Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Ketchersid, Terry L., Fresenius Medical Care North America, Waltham, Massachusetts, United States
- Maddux, Franklin W., Fresenius Medical Care North America, Waltham, Massachusetts, United States
Background
According to USRDS, over two-thirds of patients treated with hemodialysis had serum phosphate (PO4) greater than the KDIGO guidelines of 4.5 mg/dl (USRDS 2017, KDIGO 2009). In practice, most clinicians try to achieve PO4 levels between 3.0 and 5.5 mg/dl (Handbook on dialysis 2015). Proactively identifying patients who may transition from the normal clinical range to greater than 5.5 mg/dl may help clinicians target their interventions and avoid elevated PO4 levels.
Methods
We built a machine learning model with over 1000 variables to predict which HD patients will realize a PO4 lab value greater than 5.5mg/dl during the following month. We restrict our model to patients using a pharmacy specializing in renal medications with all PO4 draws less than 5.5 mg/dl in the last 180 days. Some of the variables included in this model are history of lab values, treatment vital signs, treatment no shows, comprehensive assessments from dieticians and social workers, and PO4 binder medication possession ratios. To train the XGBoost machine learning model, we utilized data from patients who were treated in Fresenius Medical Care North America clinics between January 2016 and October 2017. The results shown below are from un-seen test data with 16639 patients in 2017.
Results
For the patients enrolled in the specialty pharmacy who have all PO4 values less than 5.5 mg/dl in the previous 180 days, the prevalence of having next month’s lab draw greater than 5.5 mg/dl is approximately 10%. Results from the test data include area under the receiver operating curve (AUROC) of 0.75, sensitivity of 72% and specificity of 66%. The confusion matrix for test data is shown in Table 1.
Conclusion
We created a predictive model to help identify patients who may unexpectedly have elevated PO4 lab values in the next month. Identifying these patients may help to target interventions that mitigate negative consequences associated with high PO4 levels.
Confusion matrix for test data (n = 16639)
Actual PO4 >5.5 mg/dl | |||
Positive | Negative | ||
Prediction for PO4 >5.5 mg/dl | Positive | 7.3% | 30.6% |
Negative | 2.9% | 59.1% |
Funding
- Commercial Support –