Development of a Machine Learning Algorithm to Predict Abnormalities in Serum Phosphate in a Large Oncology Cohort.

Area of Science:

• Oncology
• Biochemistry
• Machine Learning in Healthcare

Background:

• Serum phosphate levels are critical in oncology due to their association with cancer and its treatments.
• Standardized order sets (SOS) including serum phosphate are routinely administered to oncology patients.
• Predicting phosphate abnormalities could optimize diagnostic workflows.

Purpose of the Study:

• To develop a machine learning algorithm (MLA) to predict abnormal serum phosphate levels.
• To identify if other variables within a standardized order set (SOS) can predict phosphate abnormalities.
• To assess the potential for reducing unnecessary phosphate testing in oncology patients.

Main Methods:

• An XGBoost MLA was trained using Python on 481,150 test results from 45,174 oncology patients (Jan 2019-Dec 2021).
• The model predicted abnormal phosphate (<0.5 or >1.78 mmol/L) based on other SOS variables.
• A 70%/30% train/test split was used, with subsequent validation on a separate cohort (Jan 2022-Dec 2023).

Main Results:

• The MLA achieved an area under the receiver operator curve of 0.866 on the test set.
• Implementing the model could reduce phosphate tests by over 50% (from 142,647 to 67,873), capturing 92.4% of abnormal results.
• The model demonstrated high sensitivity (0.924) and specificity (0.530), with a minimal risk of missing abnormal results (<0.1%).

Conclusions:

• A highly sensitive MLA for optimizing phosphate testing in oncology has been developed.
• Application of this MLA may lead to significant cost savings and improved healthcare efficiencies.
• The methodology is adaptable for predicting other interrelated variables within standardized order sets.