Do machine learning algorithms using routine clinical and biochemical data improve the diagnostic classification of salt-losing tubulopathies versus other causes of hypokalemia?
Machine learning algorithms using routine biochemistry can differentiate true salt-losing tubulopathies from other causes of hypokalemia with 74% accuracy, which is superior to single variables but insufficient to replace genetic testing.
Introduction: Clinically distinguishing patients with the inherited salt-losing tubulopathies (SLTs), Gitelman or Bartter syndrome (GS or BS) from other causes of hypokalemia (LK) patients is difficult, and genotyping is costly. We decided to identify clinical characteristics that differentiate SLTs from LK. Methods: A total of 66 hypokalemic patients with possible SLTs were recruited to a prospective observational cohort study at the University College London Renal Tubular Clinic, London. All patients were genotyped for pathogenic variants in genes which cause SLTs; 39 patients had pathogenic variants in genes causing SLTs. We obtained similar data sets from cohorts in Taipei and Kobe, as follows: the combined data set comprised 419 patients; 291 had genetically confirmed SLT. London and Taipei data sets were combined to train machine learning (ML) algorithms, which were then tested on the Kobe data set. Results: ) achieved a classification accuracy of 74%. This was superior to all the single biochemical variables identified previously. Conclusion: ML algorithms can differentiate true SLT in the context of a specialist clinic with some accuracy. However, based on routine biochemistry, the accuracy is insufficient to make genotyping redundant.
Wan et al. (Sat,) studied this question.