Predicting 72-Hour Fatality in Severe Hyperphosphatemia: A Comparative Analysis of Multivariate Logistic Regression...

Hyperphosphatemia is associated with several serious diseases, including chronic kidney disease, tumor lysis syndrome (TLS), rhabdomyolysis, sepsis, and acute respiratory distress syndrome. This study investigates the critical issue of predicting 72-hour fatality in patients with severe hyperphosphatemia (≥ 10 mg/dL).

We analyzed data from 530 patients treated at the Saitama Medical Center, Japan, from 2004 to 2019, including 153 72-hour fatalities. Multivariate logistic regression analysis (MLRA), Prediction One™ (Sony Network Communications Inc., Tokyo, Japan, https://predictionone.sony.biz/), and Light Gradient Boosting Machine (LightGBM) were used to predict fatalities. These methods were evaluated on a validation set of 331 patients from 2020 to 2023, including 104 fatalities. Calibration plots for training and validation data were used for comparison.

The fatality rate was 28.9% in the training data and 31.4% in the validation data. MLRA identified five fatality factors: age, low albumin, high aspartate aminotransferase, and elevated potassium and magnesium levels, with an area under the curve (AUC) of 0.848 (95% CI: 0.801, 0.890), sensitivity of 0.862, and specificity of 0.704. Prediction One™ achieved an AUC of 0.770 (95% CI: 0.722, 0.818), sensitivity of 0.654, and specificity of 0.769. LightGBM achieved an AUC of 0.948 (95% CI: 0.923, 0.973), sensitivity of 0.863, and specificity of 0.889. The validation calibration plot showed that MLRA had the closest regression coefficient to 1.0 at 0.903.

Although MLRA was the most accurate in predicting 72-hour fatalities, machine learning methods provided valuable insights into the importance of variables. Considering the high mortality rates associated with severe hyperphosphatemia, timely and accurate prognostication is essential in guiding immediate interventions and improving outcomes in emergency settings.