MLCI: Machine-Learned Comorbidity Index accepted at ICML 2026

A Machine-Learned Comorbidity Index (MLCI) by Suleman Baloch, Kishlay Jha, Alberto M. Segre, Philip M. Polgreen and Bijaya Adhikari was submitted to arXiv on 16 Jun 2026 (arXiv:2606.17450) and accepted at the 43rd International Conference on Machine Learning (ICML 2026) in Seoul, South Korea. The 35-page paper proposes a single admission-level score that "maps diagnosis codes to a single scalar" by maximizing the normalized Hilbert-Schmidt Independence Criterion, abbreviated nHSIC, against multiple clinical outcomes (paper abstract).

What is the Machine-Learned Comorbidity Index?

MLCI is an admission-level comorbidity score that learns a scalar mapping from diagnosis codes to risk by maximizing nHSIC between the learned score and several clinical outcomes. The method is explicitly designed to capture nonlinear, outcome-specific relationships rather than relying on linear, rule-based weights. The authors supply a theoretical characterization of when a unified, informative ordering across outcomes can be achieved and implement MLCI to evaluate that theory on benchmark electronic health record datasets.

How does MLCI differ from Charlson and Elixhauser?

MLCI departs from traditional comorbidity indices by addressing two limitations the paper identifies: Charlson and Elixhauser are, the authors write, "largely mortality-centric and do not align well with other clinical outcomes," and their linear, rule-based structures cannot capture nonlinear risk-outcome dependence. In contrast, MLCI optimizes a dependence criterion (nHSIC) across multiple outcomes so the learned score directly reflects outcome-specific and nonlinear relationships embedded in diagnosis codes.

What evidence do the authors provide?

The paper reports empirical results on multiple benchmark EHR datasets showing that MLCI outperforms strong baselines across multiple evaluation metrics. The authors do not publish numeric performance figures in the arXiv metadata page, but summarize that MLCI captures nonlinear risk-outcome dependence and yields superior evaluation outcomes compared with established baselines. The submission metadata includes the arXiv identifier arXiv:2606.17450 and notes the paper length as 35 pages.

Why it matters

Comorbidity scores are widely used for risk adjustment and patient stratification; shifting from mortality-centric, linear indices to a learned score that aligns with multiple clinical outcomes could change how researchers and health systems compare patients and evaluate interventions. If MLCI’s nHSIC-based objective generalizes across datasets, it would let analysts produce a single admission-level ordering that is informative for several downstream decisions instead of patching separate, outcome-specific adjustments.

What to watch

Look for the ICML 2026 presentation in Seoul and the paper’s conference materials, which will likely include the detailed numerical comparisons and code or data references that the arXiv page does not display. The next concrete signal will be the conference proceedings where the authors’ empirical tables and experimental settings should appear.