JMIR Publications

ABSTRACT

Background:

The Charlson Comorbidity Index (CCI) is widely used to quantify comorbidity burden in secondary-use clinical data. However, standard algorithms require ICD-10 diagnosis codes at greater granularity than may be available when hospital data integration centers truncate codes to three characters for privacy-preserving purposes. Because such truncation creates ambiguity across Charlson groups, frequency-informed estimation may offer a promising avenue to recover the missing granularity using national subcode frequencies.

Objective:

To develop, evaluate, and openly release five computational strategies estimating the CCI from truncated three-character ICD-10-GM codes by leveraging national four-character subcode frequencies from the German Federal Statistical Office (Destatis), and to assess how each strategy trades off encounter-level point accuracy against preservation of the cohort-level marginal CCI distribution.

Methods:

We analyzed 507,949 inpatient encounters from University Hospital Mannheim (2010–2024). Five strategies were implemented in the open-source R package miCCI including a deterministic interval estimator (S1); an inclusion-exclusion probabilistic estimator (S2); multiple imputation (S3); a Bayesian posterior median over Dirichlet-perturbed subcode probabilities (S4); and a non-negative least-squares Meta Learner combining S1–S4. Population subcode frequencies were drawn from Destatis table 23131-01. Performance against the Quan et al. ICD-10 mapping was quantified by MAE, RMSE, R², and Kolmogorov-Smirnov (KS) distance to the gold marginal, with bootstrap 95% CIs.

Results:

The two stochastic strategies achieved the strongest point accuracy (S4 MAE 0.138, 95% CI 0.137–0.140; S3 MAE 0.139, 95% CI 0.138–0.140), representing a 62% reduction versus the deterministic baseline S1 (MAE 0.361), but optimized different performance targets. S3 achieved lower squared error (RMSE 0.336 vs 0.397; R² 0.973 vs 0.962), whereas S4 better preserved the reference marginal CCI distribution (KS distance 0.021 vs 0.054; 99.23% vs 98.45% of encounters within ±1 CCI point). The Meta Learner assigned 92.4% of its weight to S3 and none to S4; it marginally improved RMSE (0.334 vs 0.336) but not MAE (0.151 vs 0.139). In an exploratory analysis across 20 ICD-10 chapters, three showed decisive differences under paired bootstrap: Neoplasms favored S3, whereas Circulatory and Digestive diseases favored S4.

Conclusions:

Frequency-informed stochastic estimation reconstructed the CCI from truncated three-character ICD-10-GM codes with MAE below 0.14 points. S3 minimized squared error, whereas S4 achieved the lowest MAE by a narrow margin and better preserved the marginal CCI distribution. The open-source miCCI package implements all strategies for three-character truncated ICD-10-GM data. Clinical Trial: Digital health; Charlson Comorbidity Index; ICD-10-GM; data truncation; multiple imputation; Bayesian inference; comorbidity scoring; administrative health data; data quality; secondary data use.