A new approach to assessing calcium status via a machine learning algorithm.

BACKGROUND AND AIMS: Calcium plays a fundamental role in biological processes. Ionized calcium (Ca2+), is the biologically active fraction, but in practice total or corrected calcium assays are routinely used to determine calcium status. MATERIALS AND METHODS: We retrospectively compared total and corrected calcium to assess the calcium status, with ionized calcium which is considered for now like the best indicator. To compensate for their lack of performance we created a machine learning algorithm to predict calcium status. RESULTS: Corrected calcium performed less well than total calcium with 58% and 74% agreement, respectively, in our population. Total calcium was especially good for hypocalcemic samples: 93% agreement versus 45% for normocalcemic and 54% for hypercalcemic samples. Corrected calcium was especially good for hypercalcemic and normocalcemic samples: 90% and 84% agreement respectively versus 40% for hypocalcemic samples. Corrected calcium is mainly faulty in hypoalbuminemia, acid-base disorders, renal insufficiency, hyperphosphatemia, or inflammatory syndrome. With our ML algorithm, we obtained 81% correct classifications. Its main advantage is that its performance are not influenced by the variables studied or the calcium status. CONCLUSION: In many situations, corrected calcium should not be used. Our ML algorithm may make a better assessment of calcium status than total calcium. Finally, if doubt, an ionized calcium assay should be performed.