In recent years, the increasing availability of real-world data from electronic health records, insurance claims, and clinical registries has opened new avenues for large-scale observational research [1]. These data sources often comprise hundreds of thousands of patient records and present opportunities to explore treatment effectiveness, patient outcomes, and health care use beyond the constraints of randomized controlled trials [2]. However, the heterogeneity and complexity of such datasets also necessitate robust methods for patient characterization, automated grouping, and risk stratification.

The Charlson Comorbidity Index (CCI) is one of the most widely used tools for quantifying a patient’s burden of comorbid diseases. Initially developed to predict 1-year mortality based on the presence of specific comorbid conditions [3], the CCI has since become a widely used metric in retrospective cohort studies, serving as a predictor of morbidity, mortality, and health care outcomes across a broad range of clinical domains [4]. In the clinical setting, its use is suggested in several clinical guidelines, for example, the 2025 European Association of Urology guideline [5]; the German S3 guidelines on prostate cancer [6], multiple myeloma [7], and diverticulitis [8]; and the National Comprehensive Cancer Network’s comprehensive geriatric assessment [9].

Over the years, many different adaptations of the original CCI have been created by adding or removing categories or changing their weights. These versions have been validated in different ways; however, they have rarely outperformed the original [4].

In 1992, Deyo et al [10] adapted the CCI for use with International Classification of Diseases (ICD), Ninth Revision, Clinical Modification (ICD-9-CM), administrative databases and modified it slightly, merging malignancy categories (solid tumors, lymphoma, and leukemia) into one, resulting in 17 instead of 19 different categories.

From that point on, a blueprint existed for using extracted administrative data to calculate a comorbidity index. Later, Sundararajan et al [11] translated the ICD-9-CM codes into the ICD, 10th Revision (ICD-10), Australian Modification. A year later, Quan et al [12] revised the codes again, creating a new ICD-10 code list for every comorbidity.

Quan et al [12] intended to keep their mapping country independent by restricting the ICD codes to 4 characters (ie, the first letter and 3 digits) [12]. However, the list of ICD codes has since undergone substantial revisions, introducing new subcodes and modifying disease definitions, with many countries developing their own versions or modifications, such as the ICD-10, German Modification (ICD-10-GM). These refinements have been made predominantly at the most granular level (the fourth digit or the fifth character of a code), making country-specific mappings necessary—at least theoretically.

Despite these changes, the mapping published by Quan et al [12] in 2005 remains one of the most popular approaches and continues to be frequently cited in today’s research (approximately 10,000 citations on Web of Science as of January 29, 2026) while also acting as the basis of automated CCI calculations in popular Python or R packages [13,14]. To our knowledge, there is currently no openly available mapping that has been developed specifically for German administrative data. Therefore, the aim of our study was to develop an open-source tool enabling reproducible comorbidity adjustment for German real-world data. We benchmarked our approach against the mappings by Quan et al [12] in a large cohort of 515,827 inpatient cases from a German tertiary care center (2010-2024) using year-specific mappings for each case and demonstrated that our mappings are better aligned with current German coding practice.

To address this potential gap between the current German coding practices and the available mapping, we developed new mappings for calculating the CCI using the ICD-10-GM for the years 2004 to 2026. The goal of this study was to (1) create updated, country-specific, year-specific CCI mappings ensuring compatibility with current German coding standards; (2) compare the newly developed mappings with the widely available mapping by Quan et al [12] from 2005 evaluating agreement and category-level differences, thus investigating whether the new mapping is the optimal way to achieve the above-mentioned compatibility; and (3) provide an easy-to-use, transparent, open-source tool using the mapping, thus facilitating the optimal comorbidity classification for German patient data and future adaptations.

We aimed to improve the compliance of the automated comorbidity classification and scoring with actual clinical practice in German hospitals. This will enable more adequate stratification of the patients in retrospective clinical research.

As the goal of this project was not the revalidation of clinimetric properties of the CCI but, rather, updating the currently used mapping and keeping it compatible with other country-specific CCI adaptations, the very same categories and weights were used as in the version by Quan et al [12].

As a reference, we used the ICD-10 mapping as published by Quan et al [12] in 2005. We then constructed a single new mapping tailored to the 2024 ICD-10-GM using a prespecified, multistep process as follows.

First, we adopted the adaptation by Deyo et al [10] of the CCI (17 categories instead of 19), effectively setting the same framework for our mapping as Quan et al [12] did in 2005.

Second, 2 clinicians (FS and PPS; FS also being a coding expert with a decade of experience) independently screened the 2024 ICD-10-GM list as provided online by the German Federal Institute for Drugs and Medical Devices (Bundesinstitut für Arzneimittel und Medizinprodukte; BfArM) to identify candidate codes for every category. The selection process was based first on the clinical term defining the category and on the descriptions of the comorbid categories from the 1987 paper by Charlson et al [3].

Third, the independently created code sets were compared. Discrepancies were enumerated and resolved in a consensus meeting. For each discrepant code, discussion was held until a consensus was achieved. In cases in which no consensus was achieved, a disease-specific coding expert was consulted.

The 2024 iteration was developed first because it corresponded to the most recent year available in our patient data.

After establishing the 2024 mapping, we audited each ICD-10-GM annual release (2004-2026) and compiled a year-specific list of added, removed, split, merged, or redefined codes for each year. These changes were then independently revised by the same 2 clinicians, and the proposed changes in the mappings for each year underwent the same multistep process explained above. The resulting mappings were intended to be applied according to each case’s discharge year.

A new feature as compared to the mapping by Quan et al [12] was the introduction of Boolean rules in some of the categories, which enabled more specific scoring (eg, code A and code B score points when they are both present but not separately).

To summarize, the new mappings included new sets of ICD codes chosen according to the above-mentioned algorithm. For each year, a new mapping was created, which enabled coding up to the most granular level, unlike the mapping by Quan et al [12], where the limitation of 4 characters per code was introduced with the intention of the single mapping being transferable to subsequent years and other countries’ coding practices. The next level of granularity was achieved by the introduction of Boolean rules so that some codes could only score points in specific combinations according to German coding guidelines.

To practically use the mapping and calculate the CCI, we created a Python library with one core function requiring one or more ICD codes and a mapping definition as input and returning the score (number of points) and the scored categories.

The mapping definition was intentionally separated from the CCI scoring logic to enable easy modifications without having to change the central component of the algorithm.

The library consists of 24 separate mappings: a mapping described and validated beforehand by Quan et al [12] in an unchanged form and the 23 mappings built by authors FS and PPS based on the ICD-10-GM for years 2004 to 2026.

The algorithm was then applied to a large dataset of German patients’ administrative data.

The results obtained by applying our year-specific mappings—for each case, the mapping for the corresponding year was applied—were pooled and compared with the results obtained using the mapping by Quan et al [12], which stayed the same throughout all years.

The mappings were developed—as mentioned above—for years 2004 to 2026. However, as our study population included records from the years 2010 to 2024 (see the Study Population and Data Source section), only the mappings for those years were deployed.

The statistical analyses were conducted in Python (version 3.12.3; Python Software Foundation) [15] and R (version 4.4.2; R Foundation for Statistical Computing) [16]. For each case, we computed 2 total CCI scores: one using the year-specific ICD-10-GM mapping (selected by discharge year) and one using the mapping by Quan et al [12].

We first calculated the proportion of cases with different total CCI scores between the 2 approaches. To quantify the agreement, the intraclass correlation coefficient (ICC) was calculated. We used a 2-way mixed-effects model for absolute agreement based on single measurement because we wanted to assess only the absolute agreement of the scores (and not their consistency) between the 2 (fixed) alternative mappings [17]. The dependent variables in this analysis were the CCI scores per case. The analysis was conducted using the irr package [18] in R. We reported the resulting ICC and P value.

For each of the 17 categories, we applied the McNemar test to the discordant pairs. When the number of discordant pairs was less than 25, we used the exact binomial test; otherwise, continuity-corrected chi-square approximation was used. Tests were 2 sided, with an α of .05. Given the explorative nature of the analysis, no adjustment for multiplicity was applied. The calculation of the McNemar test was performed using the statsmodels library (version 0.15.0) [19] in Python.

Category-level concordance was displayed as a diverging bar chart ordered by total discordant count. Figures were produced using the matplotlib package (version 3.5.3) [20] in Python. Further investigation was undertaken when a significant difference was exhibited to determine whether the difference was the result of erroneous or better representation of ICD codes by one of the approaches. For the category with most total discrepancies, an additional bar chart was created to show its proportion in different diverging scores.

To assess the score-level impact of mapping discordance, we calculated ΔCCI for each case as the difference between the CCI obtained by our mapping and by the mapping by Quan et al [12]. We plotted the distribution of nonzero ΔCCI values.

The code list with descriptions (ICD-10-GM 2004-2026) was obtained from the official website of the BfArM [21].

The study was approved by the Ethics Committee II, Medical Faculty Mannheim, Heidelberg University (approval ID 2024-885). The requirement for informed consent was waived due to the retrospective analysis of pseudonymized data.

Table 1 shows a comparison of the mapping by Quan et al [12] from 2005 and our mapping based on the ICD-10-GM from 2024 (developed using the process detailed in the Methods section). For clarity purposes, the other year-specific mappings are not presented in this table but are openly available online (see the Data Availability section).

We created year-specific ICD-10-GM mappings (2004-2026) via dual independent review with consensus and implemented them in an open-source Python library. The new mappings not only included corrected alignment of codes to the comorbidity categories but also introduced Boolean rules that allowed for an even more precise algorithm.

Applied to 515,827 inpatient cases from the years 2010 to 2024, the new mappings showed excellent agreement with the mapping by Quan et al [12] (ICC=0.992; P<.001), whereas 3.1% (16,136/515,827) of cases received different total CCI scores.

Discordance was highly concentrated in a single administrative code—Z95.81 (presence of an implanted port system)—which was present in almost half (7896/16,136, 48.9%) of all discordant cases; because this code is nondiagnostic for peripheral vascular disease, it was excluded in our mappings.

Other codes were excluded because their BfArM [21] definitions clearly indicated that they were not diagnostic of the respective comorbidity category.

By applying those changes, and after inspecting individual discordant codes, we conclude that our new, year-specific mappings more precisely reflect the coding reality in the German health care system.

We find it important to discuss some of the contents of Table 3 and the exclusion of selected codes that, in our opinion, are not diagnostic for the categories.

The most abundant discordant code, Z95.81, stands for the presence of a surgically implanted vascular catheter system (eg, Broviac catheter or port system). As the presence of a vascular access device does not, by itself, indicate peripheral vascular disease, this finding underscores the necessity of adapting automated comorbidity scoring tools to reflect contemporary coding definitions.

Codes starting with I85 denote esophageal varices. For codes I85.0, I85.9, and I86.4, the code definitions by BfArM [22] state clearly that they are not to be used if the underlying condition is liver disease. The coding guidelines designate combination codes I98.2 and I98.3 for these conditions, which were included in our definition via Boolean operators. In contrast, the ICD-10, Clinical Modification (ICD-10-CM), introduced codes for secondary esophageal varices as I85.10 and I85.11 in 2007 [23]. The definition by Quan et al [12] would include primary esophageal varices that are not due to underlying liver diseases both in current German and US coding standards.

Code I73.9 is described as “peripheral vascular disease” and was not scored as such in our mapping. This is because the code does not include claudication as a consequence of occlusion disease but, rather, describes functional causes such as arterial spasm [24].

Codes K74.71 and K74.72 were not scored in the category “mild liver disease” using our mapping but in “severe liver disease” as they describe advanced stages of liver cirrhosis [25].

K55.8 and its subcodes were excluded because, according to the ICD-10-GM classification, ischemic intestinal vascular diseases are specifically represented by K55.0 (“acute vascular disorders of intestine”) and K55.1 (“chronic vascular disorders of intestine”) [26]. K55.8 is reserved for other specified vascular diseases of the intestine not captured by these ischemic categories and was not considered equivalent to intestinal ischemia for the purpose of this mapping.

Z95.80, the presence of a cardiac assist system, does not unambiguously imply the condition of the peripheral vessels, which is why this code was excluded from the “peripheral vascular disease” category. Similarly, the current version of the ICD-10-CM includes code Z95.810, which signifies the “presence of automatic (implantable) cardiac defibrillator” [27], which would also not be diagnostic for the comorbidity.

Cauda equina syndrome—codes G83.40, G83.41, and G83.49—does not inherently imply paraplegia or hemiplegia, which is why we removed it.

The version of the CCI that we chose was the modification by Deyo et al [10], not the original one presented by Charlson et al [3] in 1987. We decided to use this modification because of the already available ICD-10 mappings and their modifications, such as the one by Quan et al [12]. However, the original index with 2 additional comorbidity categories could deliver better, more precise results. This would require further investigation. Our transparent tooling could readily accommodate this in future analyses.

It is important to note that even a near-perfect algorithm and mappings are still dependent on the quality of clinical documentation and ICD coding practices. A Canadian 20-year validation of ICD-10 administrative data using the CCI found that several comorbidities were increasingly undercoded over time [28]. Mapping accuracy cannot offset systematic undercoding, but errors compound. Each stage—documentation, coding, and mapping—matters for the validity and quality of retrospective observational research.

Even though modern statistical methods, often supported by—but not limited to—machine learning algorithms, can yield very cohort-specific comorbidity indexes with excellent local calibration, this methodology sacrifices comparability across studies and settings. General indexes such as the CCI enable risk adjustment and cross-study synthesis in national systems outside Europe, as shown in the Korean claim-based validation and scoping review by Shin and Han [29]. We therefore view tailored models as complementary to, not replacements for, standardized indexes such as the CCI.

We provide year-specific, versioned mappings that enable programmatic calculation of the CCI using German administrative database codes (ICD-10-GM). In a direct comparison with the widely used mapping by Quan et al [12], agreement was very high.

Inspection of the discordant codes suggests that the discrepancies reflect clinically implausible inclusions in the older mapping when applied to current ICD-10-GM definitions.

The obtained results demonstrate that country- and year-specific mappings are necessary to avoid misclassification when legacy mappings are applied to contemporary German coding practice and that a small set of high-frequency administrative codes can substantially bias automated scoring at scale. The released, versioned mappings and lightweight Python implementation enable reproducible, year-appropriate CCI scoring. Our work provides valuable practical tools for CCI scoring when using German administrative data.

The calculation of comorbidity indexes from ICD codes is not novel per se, with the first publications dating back to 1992 [10]. The novelty of our work lies in updating the mapping to reflect the country- and year-specific ICD-10 codes used in the German health system and making it easily accessible, interoperable, reusable, and adaptable for future modifications.

This study quantifies, to our knowledge for the first time, the magnitude and origin of mapping-related misclassification risk when outdated mappings are applied to current coding practices—demonstrating that maintaining country- and year-specific mappings is critical for valid risk adjustment.

The mappings are designed to be regularly updated just as the ICD-10 iterations are released every year. The implemented architecture, which stores the code lists in separate, open-format files, intentionally supports these updates. It also allows for the addition of new comorbidity scores (eg, that by Elixhauser or the original CCI).

Performance is bounded by and dependent on the coding quality of the processed data. The present work targeted the ICD-10-GM for the years 2004 to 2026, the implication being that it requires yearly maintenance and alignment to match the changes in coding practices as the German ICD-10 list is updated every year.

For CCI scoring in German datasets, locale- and version-specific mappings should be preferred over international lists. The case of a single code being present in almost 50% of discordances (Z95.81) demonstrated that a single change may have significant implications for overall scoring when processing data programmatically. This impact might even be increased in specialized patient subcohorts.

When reporting comorbidity adjustment, the CCI variant and the applied mapping should be stated. In studies using ICD-coded data, implementations often follow the mapping by Deyo et al [10] via the ICD mapping by Quan et al [12], yet this is frequently left implicit and only becomes clear through the cited works. We therefore recommend clearly documenting (1) the CCI specification used (original 19-category CCI vs 17-category modification by Deyo et al [10]) and (2) the mapping source and its version to ensure reproducibility and comparability across studies, for example, “Comorbidities were assessed using the original 19-item Charlson Comorbidity Index; scoring was performed by manual chart review” or “Comorbidities were assessed using the 17-item Deyo modification of the Charlson Comorbidity Index; scoring was performed automatically from ICD-10-CM codes using Quan et al’s mapping from 2005.”