The use of international classification of diseases codes to identify hospital admissions linked with adverse drug events: Validation study

Abstract

Aims

Several methods exist to identify hospital admissions related to adverse drug events (ADEs). Clinical adjudication by healthcare professionals is the gold standard but is labour‐intensive. Spontaneous reporting and routinely collected healthcare data using a set of International Classification of Diseases (ICD) codes often underestimate the prevalence of ADE‐related admissions. Expanding the set of ICD codes could improve detection; however, validation is limited. The objective was to describe the agreement between ADE‐related ICD‐10 codes and clinically adjudicated ADE‐related admissions in 2 settings.

Methods

This study analysed 2 datasets: 1102 readmissions from a hospital in the Netherlands (180 ADE‐related) and 1228 admissions from a hospital in the Czech Republic (195 ADE‐related). Clinical adjudication involved expert review including causality assessment to identify ADE‐related hospital admissions. The sensitivities and specificities were calculated for a narrow code set (higher drug‐likelihood codes containing words like drug‐induced) and a broad code set of ICD‐10 codes (including codes very likely, likely and possibly ADE‐related).

Results

The narrow ICD‐10 set showed a sensitivity of 3% (95% confidence interval [CI] 2–6%) and a specificity of 99.6% (95% CI 99–100%). The broad set increased sensitivity to 27% (95% CI 23–32%), with specificity decreasing slightly to 92% (95% CI 91–94%). Preventable ADEs were identified less frequently with both ICD‐10 code sets.

Conclusions

Only 3% of ADE‐related admissions were detected by the narrow ICD‐code set and 27% by the broad code set without a significant drop in the specificity. ADE‐related ICD codes seem to serve as triggers for 1 in 4 ADE‐related hospital admissions.

What is already known about this subject

• Adverse drug events (ADEs) are a significant and preventable cause of hospitalizations, but spontaneous reporting results in substantial underreporting.

• International Classification of Diseases (ICD) code‐based methods improve detection but have limited sensitivity, underestimating the prevalence of ADE‐related hospital admissions.

• Broadening ICD code sets shows promise for ADE detection but needs validation across diverse healthcare settings.

What this study adds

• Clinically validated ICD code sets can be used for ADE identification in order to guide preventive strategies.

• Narrow ICD code sets (restricted to terms such as drug‐induced) showed poor agreement with clinically adjudicated ADE‐related admissions, detecting only 3% of cases.

• Broadening the ICD code set increased sensitivity to 27%, improving ADE detection without significantly compromising specificity.

1. INTRODUCTION

Adverse drug events (ADEs) represent harm resulting from medication use. The term ADE encompasses events that result not only from the appropriate use of medications, e.g., adverse drug reactions (ADRs) but also from the inappropriate use of medications, i.e., ADEs due to medication errors. ¹ ADE‐related hospital admissions and readmissions pose a significant and often avoidable burden on healthcare systems. The 2022 report by the Organization for Economic Cooperation and Development highlighted that medication‐related harm accounts for 1 in 10 hospitalizations in Organization for Economic Cooperation and Development countries. ²

ADEs that result in hospital admissions or the prolongation of hospital admissions represent a serious type of ADE. ³ As a result, healthcare professionals are required to report any suspected cases of hospital admissions that arise not only from ADRs but also from medication errors. ⁴ , ⁵ Spontaneous reporting has proven insufficient in capturing ADE‐related hospital admissions, resulting in significant underreporting across various countries and settings. ⁶ , ⁷ Additionally, clinical adjudication of ADEs through medical record reviews, including causality assessment, regarded as the gold standard for ADE detection, is not feasible on a large scale due to the high volume of admissions. ⁸

To overcome the underreporting of medication‐related hospital admissions via spontaneous reporting, various methods of active surveillance are used to augment this reporting. Several studies have used International Classification of Diseases (ICD) codes to identify medication‐related admissions in comprehensive national administrative databases. ⁹ However, the diversity and variability of ICD codes used in the literature pose challenges in accurately determining the prevalence of medication‐related hospital admissions. ¹⁰ Only a limited number of studies have evaluated the sensitivity and specificity of ICD code sets in identifying clinically adjudicated ADEs, ¹¹ indicating relatively low sensitivity and highlighting the limitations of using routinely collected ICD‐coded diagnoses to detect ADEs. Several studies have suggested that the use of obvious medication‐related ICD codes (e.g., with drug‐induced terms such as drug‐induced hypoglycaemia without coma) underestimates the true prevalence of ADEs. ¹² , ¹³ , ¹⁴ , ¹⁵

To address these challenges, Hohl et al. ¹¹ synthesized a broad set of ICD codes that includes not only medication‐related ICD codes but also indicators of very likely, likely or possible ADEs. Unlike earlier methods that depended on clinician recognition of ADEs, the broader set of ICD codes synthesized by Hohl et al. provides a more comprehensive approach to capturing potential ADEs and demonstrates higher sensitivity than previously used ICD code sets, with sensitivity increasing from 6.8 to 28.1%. ¹⁶ Meanwhile, specificity decreased only slightly from 99 to 87.7%, ¹⁶ indicating more effective detection of ADEs.

Despite its promising performance, the broad ICD code set has not been evaluated in different settings. Therefore, its reliability and applicability beyond the Canadian setting ¹⁶ , ¹⁷ are yet to be determined. Moreover, no study has estimated the sensitivity of this ICD code set in detecting potentially preventable ADEs. This is particularly important because preventable ADEs can be mitigated with appropriate interventions (e.g., medication review). The objective of this study was to describe the ICD code set's sensitivity and specificity to detect ADE‐related hospital admissions in 2 different countries (the Netherlands and the Czech Republic). This study described the agreement between ADE‐related ICD‐10 codes and clinically adjudicated ADE‐related admissions by calculating the sensitivity and specificity of different ICD‐10 code sets synthesized by the systematic review of Hohl et al. ¹¹ In addition, we aimed to compare the sensitivity and specificity of ICD code sets for identifying potentially preventable vs. nonpreventable ADE‐related hospital admissions.

2. METHODS

2.1. Study setting

This study used datasets from 2 different locations: University Hospital Hradec Králové in Hradec Králové, Czech Republic and OLVG Hospital in Amsterdam, the Netherlands. Both the Czech Republic and the Netherlands have universal healthcare systems based on compulsory coverage.

2.2. Eligibility criteria and study population

We included hospital admissions (dataset from the Czech Republic) and hospital readmissions within 30 days after discharge (dataset from the Netherlands) for adult patients. These datasets were obtained from previously conducted studies ¹⁸ , ¹⁹ (Table 1). Paediatric patients (age <18 years) and hospital (re)admissions with missing ICD‐10 codes were excluded from the analysis.

TABLE 1.

Characteristics of study datasets.

Characteristic	Dataset from the Czech Republic	Dataset from the Netherlands
Study	Očovská et al. 2022	Uitvlugt et al. 2021
City	Hradec Králové, Czech Republic	Amsterdam, the Netherlands
Setting	University Hospital Hradec Králové	OLVG Hospital
Number of included adult hospital (re)admissions	1228 unplanned adult admissions	1102 unplanned adult readmissions
Age (years), mean (SD)	68 (13)	64 (17)
Sex	55% male	50% male
Departments	Admissions via the Department of Emergency Medicine to any department in the hospital, primarily internal medicine (including cardiology, gastroenterology, metabolic care, gerontology, haematology), surgery, neurology and pulmonology	Cardiology, gastroenterology, internal medicine, neurology, psychiatry, pulmonology and general surgery
Time period	August 2018–November 2018	July 2016–February 2018
The most common ICD diagnoses at admissions	I50.0 Congestive heart failure (6%) I63.0 Cerebral infarction due to thrombosis of precerebral arteries (3%) I21.4 Non‐ST elevation (NSTEMI) myocardial infarction (3%)	T81.4 Infection following a procedure (4%) C34.9 Malignant neoplasm of unspecified part of bronchus or lung (3%) J18.9 Pneumonia, unspecified organism (3%)
Clinical adjudication	Method: OPERAM adjudication guide Clinical adjudication was performed by a pharmacist, with validation conducted by board‐certified clinical pharmacists.	Method: Kramer algorithm Clinical adjudication was performed by residents of the different departments (n = 7) and hospital pharmacy resident. Validation was performed by internal medicine physician and hospital pharmacist/clinical pharmacologist.

Abbreviations: ADE, adverse drug event; ICD, International Classification of Diseases; SD, standard deviation.

The original study from the Czech Republic ¹⁸ included 1252 hospital admissions, but after excluding paediatric cases, 1228 admissions remained for analysis. These admissions occurred across various hospital departments from August to November 2018.

The original study from the Netherlands ¹⁹ included 1111 hospital readmissions, but after removing those with missing ICD codes, 1102 readmissions were included. These readmissions also occurred across various hospital departments, covering the period from July 2016 to February 2018.

2.3. Clinical adjudication of ADEs

The gold standard for identifying ADEs was clinical adjudication using medical record review. Clinical adjudication was performed by pharmacists and physicians and included an assessment of causality, which considered alternative causes, plausible time relationships and the assessment of the ADE's contribution to hospital (re)admissions. Only ADEs that were the main reason or contributed to hospital (re)admissions were included. ADEs resulting from medication errors were classified as preventable.

2.4. ICD‐10 codes

An ADE‐related ICD‐10 code was defined as an ICD code that was listed by the systematic review conducted by Hohl et al. ¹¹ to identify ADEs in administrative health data (see the electronic supplementary material of Hohl et al. ¹⁶ ). These codes were classified into drug‐likelihood categories, which are described in Table 2. Hohl et al. adapted the drug‐related likelihood categories defined by Stausberg and Hasford, ²⁰ who classified ICD codes into 7 categories (A1, A2, B1, B2, C, D, E), each reflecting its validity as an indicator for an ADE. Codes in categories A and B explicitly indicate drug‐related causation in the ICD diagnosis, whereas codes in categories C, D and E lack an explicit reference to a medication.

TABLE 2.

Drug‐related likelihood categories by Hohl et al.

Drug‐related likelihood categories		Example	ICD code set
A1	ICD‐10 code description includes the phrase induced by medication/drug	E16.0: Drug‐induced hypoglycaemia without coma	Both narrow and broad set
A2	ICD‐10 code description includes the phrase induced by medication or other causes	I42.7: Cardiomyopathy due to drugs and other external agents
B1	ICD‐10 code description includes the phrase poisoning by medication	T40.2: Poisoning by other opioids (includes morphine and codeine)
B2	ICD‐10 code description includes the phrase poisoning by or harmful use of medication or other causes	T50.9: Poisoning: Other and unspecified drugs, medicaments and biological substances
C	ADE deemed very likely, although the ICD‐10 code description does not refer to a drug	K71.9: Toxic liver disease, unspecified	Broad code set only
D	ADE deemed likely, although the ICD‐10 code description does not refer to a drug	K25.0: Gastric ulcer: Acute with haemorrhage
E	ADE deemed possible, although the ICD‐10 code description does not refer to a drug	E87.1: Hypo‐osmolality and hyponatraemia

Abbreviations: ADE, adverse drug event; ICD, International Classification of Diseases.

The narrow code set included higher drug‐likelihood codes (e.g., containing the words drug‐induced or due to drugs) that corresponded with drug‐likelihood categories A and B. The broad code set included lower drug‐likelihood codes (very likely, likely, possibly ADE‐related) as well, i.e., categories A–E.

2.5. Outcome measures and data analysis

The primary outcomes of this study included sensitivities and specificities for the narrow and broad ADE‐related ICD codes. ¹¹

A true positive was defined as the presence of an ICD code from either the narrow or broad set for a patient who had also been identified as having an ADE through clinical adjudication. A false positive was defined as the presence of an ICD‐10 code from the code set that was not clinically adjudicated by healthcare professionals as an ADE.

The sensitivity of the ICD code set was calculated as the proportion of true positive cases to the total number of clinically adjudicated ADE‐related admissions (true positives divided by the sum of true positives and false negatives).

The specificity of the ICD code set was calculated by dividing the number of true negatives by the total number of admissions that were not clinically adjudicated as ADEs (true negatives divided by the sum of true negatives and false positives).

We also calculated the positive predictive value (PPV) and negative predictive value (NPV). Unlike sensitivity and specificity, PPV and NPV are dependent on the prevalence of the condition (ADE‐related admission) in the population. The PPV of the ICD code set was calculated by dividing the number of true positives by the number of admissions flagged by ICD codes as ADEs (true positives divided by the sum of true and false positives). The NPV of the ICD code set was calculated by dividing the number of true negatives by the total number of admissions not flagged by ICD codes as ADEs (true negatives divided by the sum of true and false negatives).

The subgroup analysis focused on the preventability of ADE‐related admissions, allowing us to determine whether differences exist in ADE detection through ICD diagnoses in relation to preventability.

The sensitivity and specificity of the ADE‐related ICD code set were also calculated separately for each drug‐related likelihood category as defined by Hohl et al., ¹¹ see Table 2.

The data were analysed using descriptive statistics to summarize key characteristics. The calculation of confidence intervals was based on a 95% confidence level to provide an estimate of uncertainty around the sensitivity, specificity, PPV and NPV measures.

3. RESULTS

3.1. Study flow chart

Table 1 provides a summary of the characteristics of the study datasets and Figure 1 displays the study flow chart. The dataset from the Czech Republic included 1228 admissions, while the dataset from the Netherlands included 1102 readmissions. Within the dataset from the Czech Republic, 195 hospital admissions (15.9%) were classified as ADE‐related. In the dataset from the Netherlands, 180 hospital readmissions (16.3%) were classified as ADE‐related. In the Czech dataset, there was a total of 100 preventable ADE‐related admissions (51% of all ADE‐related admissions), while in the dataset from the Netherlands, there was a total of 72 preventable ADE‐related admissions (40% of all ADE‐related admissions).

FIGURE 1.

Study flow chart. ADE, adverse drug event.

3.2. Sensitivity and specificity of the ICD code sets

The sensitivities and specificities of ICD code sets to detect ADE‐related admissions are presented in Table 3. The narrow ADE‐related ICD code set identified similar proportions of ADE‐related admissions: 3.1% (95% confidence interval [CI]: 0.7–5.5%) in the dataset from the Czech Republic and 3.3% (95% CI: 0.7–6.0%) for the Dutch dataset. The broad ADE‐related ICD code set identified a higher proportion of ADE‐related admissions; 27.2% (95% CI: 20.9–33.4%) for the Czech Republic dataset and 27.2% (95% CI: 20.7–33.7%) for the Dutch dataset.

TABLE 3.

Sensitivity, specificity, PPV and NPV of ICD code sets to detect ADE‐related admissions.

ICD code set	Identification with ICD codes	Dataset from the Netherlands			Dataset from the Czech Republic
		ADE	No ADE	Total	ADE	No ADE	Total
Narrow ICD code set	Identified	6	4	10	6	4	10
	Not identified	174	918	1092	189	1029	1218
	Total	180	922	1102	195	1033	1228
	Sensitivity	3.3% (95% CI: 0.7–6.0%)			3.1% (95% CI: 0.7–5.5%)
	Specificity	99.6% (95% CI: 99.1–100.0%)			99.6% (95% CI: 99.2–100.0%)
	PPV	60.0% (95% CI: 29.6–90.4%)			60.0% (95% CI: 29.6–90.4%)
	NPV	84.1% (95% CI: 81.9–86.2%)			84.5% (95% CI: 82.4–86.5%)
Broad ICD code set	Identified	49	63	112	53	85	138
	Not identified	131	859	990	142	948	1090
	Total	180	922	1102	195	1033	1228
	Sensitivity	27.2% (95% CI: 20.7–33.7%)			27.2% (95% CI: 20.9–33.4%)
	Specificity	93.2% (95% CI: 91.5–94.8%)			91.8% (95% CI: 90.0–93.4%)
	PPV	43.8% (95% CI: 34.6–52.9%)			38.4% (95% CI: 30.3–46.5%)
	NPV	86.8% (95% CI: 84.7–88.9%)			87.0% (95% CI: 85.0–89.0%)

Abbreviations: ADE, adverse drug event; CI, confidence interval; ICD, International Classification of Diseases; PPV, positive predictive value; NPV, negative predictive value.

The list of true positive ICD codes can be found in Tables S1 and S2.

Table S3 presents the subgroup analysis regarding the preventability of ADE‐related admissions. The subgroup analyses highlighted the lower sensitivity of both the narrow and broad ICD code sets to detect preventable ADEs compared to nonpreventable ADEs, although differences were small. The sensitivity for preventable ADEs was 25%, while for nonpreventable ADEs it was 29–30%. Specificity remained high, exceeding 90%.

The subgroup analysis for each category of drug‐related likelihood can be found in Table 4. This subgroup analysis revealed that the drug‐related likelihood category with the highest sensitivity to detect clinically adjudicated ADE was category E (ADE deemed possible with sensitivity over 15% and specificity over 90%).

TABLE 4.

Sensitivity and specificity grouped by drug‐related likelihood categories.

		Sensitivity		Specificity
Drug‐related likelihood categories of ICD codes		Dutch	Czech	Dutch	Czech
A.1	Induced by medication	1.7%	2.1%	99.7%	100%
A.2	Induced by medication or other causes	0.6%	0.5%	99.9%	99.9%
B.1	Poisoning by medication	1.1%	0.5%	100%	99.8%
B.2	Poisoning by or harmful use of medication or other causes	‐	‐	100%	99.9%
C	ADE deemed very likely	2.8%	1.5%	99.0%	100%
D	ADE deemed likely	5.6%	6.7%	99.3%	99.2%
E	ADE deemed possible	15.6%	15.9%	95.2%	92.9%

Abbreviations: ADE, adverse drug event; ICD, International Classification of Diseases.

The denominator in the calculation of sensitivity: n = 195 (dataset from the Czech Republic), n = 180 (dataset from the Netherlands)—number of clinically adjudicated ADE‐related admissions (the sum of true positives and false negatives).

The denominator in the calculation of specificity: n = 1033 (dataset from the Czech Republic), n = 922 (dataset from the Netherlands)—number of non‐ADE‐related admissions (the sum of true negatives and false positives).

4. DISCUSSION

4.1. Summary of the findings

The study revealed limited agreement between ADE‐related ICD‐10 codes and clinically adjudicated ADE‐related admissions, with only 3% of cases identified using a narrow ICD code set. This indicates that relying solely on a narrow code set is insufficient for capturing ADE‐related hospital admissions. Nonetheless, the high specificity observed indicates that while the narrow set is selective, it is unlikely to falsely identify non‐ADE cases. Broadening the code set led to an 8 to 9‐fold increase in sensitivity, capturing more ADE‐related admissions without a significant drop in specificity. The findings highlight the potential of using a broad set of ICD codes as a more accurate approach for identifying ADE‐related hospital admissions.

Although the sensitivity of the broad ICD code set for detecting preventable ADEs was lower than for nonpreventable ADEs (25 vs. 29–30%), it still represents a significant improvement over traditional detection methods in daily practice, such as spontaneous reporting or the narrow ICD code set.

4.2. Comparison with other studies

The systematic review by Hohl et al. ¹¹ found that only 2 ¹⁶ , ²¹ out of 41 studies using ICD codes to detect ADEs reported both sensitivity and specificity for their ICD‐10 code sets. Since then, only a few studies have evaluated these metrics, with sensitivity ranging from 3.4 to 17.8% and specificity from 92.2 to 99.7%. ¹² , ¹⁶ , ¹⁷ , ²¹ , ²² In the Canadian study by Hohl et al., ¹⁶ using a broader ICD code set increased sensitivity from 6.8 to 28.1% while decreasing specificity from 99 to 87.7%. Our results closely aligned with Hohl et al.’s findings, showing an increased sensitivity from 3 to 27.2% without a corresponding drop in specificity. In contrast, a recent study from Canada by Wickham et al. ¹⁷ observed significant variations in sensitivity and specificity depending on the dataset used (emergency department vs. inpatient data). Nevertheless, broadening the ICD code set consistently enhanced sensitivity by 3–4 times.

4.3. ICD codes vs. medical record review

The prevalence of ADE‐related hospital admissions varies based on the detection method used. Clinical adjudication of medical records represents the gold standard for ADE detection, while ICD codes tend to underestimate the prevalence of medication‐related hospital admissions. ⁸ , ¹² , ¹⁴ , ¹⁵ , ²³ , ²⁴ For example, a study from Slovenia ²⁴ found that only 0.2% of medication‐related admissions were identified by ICD codes. Relying solely on these codes underestimates the prevalence of ADEs as a cause of hospital admission. ¹² , ¹⁴ The detection of ADEs using narrow ICD code sets depends on clinicians accurately identifying and coding these events, but this is often done suboptimally, resulting in ADEs being coded as diseases without medication references. ²² , ²⁵ Conversely, broadening the ICD code set by including diagnostic codes unrelated to medications requires additional effort to confirm ADEs through detailed medical record reviews. ¹³ , ²⁶ The complex process of coding is complicated by changing administrative coding rules, impacting accurate ADE detection. ²⁶ Financial incentives, such as diagnosis‐related group creep, where hospitals code conditions with higher reimbursement rates, might drive variation in coding practices in the hospital setting. ²⁷ Incomplete coding of ADEs can also result from limited slots for secondary diagnoses ²⁷ or discrepancies between electronic and paper medical records. ²⁵

The standard use of ICD coding in hospitals makes this method both available and cost‐efficient. ²² , ²⁸ , ²⁹ The strength of the ICD code methodology lies in its ability to detect rare ADEs and its simple integration with large‐scale population datasets. The primary limitation of ICD codes is their inability to independently identify ADEs, necessitating their use alongside other detection methods. ¹⁵

Studies combining both methods of detection have shown that ICD codes and medical record reviews identify different types of ADRs. ⁸ , ¹² , ¹⁷ For instance, Nair et al. ¹² found that haematological (haemorrhage, agranulocytosis) and metabolic ADRs (hyponatraemia, hyperkalaemia) were frequently identified through ICD coding, while cardiovascular ADRs were detected only through medical record review. A study from the UK ¹⁵ found ICD codes more sensitive for oncology ADRs, while a recent study from Canada ¹⁷ showed a higher sensitivity for severe ADEs such as arrhythmias. These variations highlight the need for further research into the differences in ADE detection methods.

4.4. Significance and possible applications

In 2012, the European pharmacovigilance legislation ³ expanded its focus to cover noxious and unintended effects resulting from medication use outside the terms of the marketing authorization, such as overdose, misuse, abuse and medication errors. As a result, healthcare professionals are required to report medication‐related hospital admissions that arise not only from ADRs but also from medication errors, i.e., ADEs. Given the focus on ADE‐related admissions in pharmacovigilance and the prevalent underreporting by healthcare professionals, there is an urgent need for new methods of ADE detection. Continuous use of database methodologies is recommended to complement the system of spontaneous reporting. ⁸

The information on the specificity and sensitivity of ADE‐related ICD‐code sets and specific ICD codes provides insights into potential applications. There is an inherent trade‐off between sensitivity and specificity. A high specificity is achieved at the expense of lower sensitivity, which substantially restricts the clinical applicability of the narrow ICD code set detection method. Conversely, the broader ICD code set offers higher sensitivity but at the cost of reduced specificity, making confirmatory medical record reviews necessary. As a result, the ICD codes with excellent specificity (100%) to detect ADE‐related admissions could be directly reported to national pharmacovigilance centers and used in pharmacoepidemiology research. Meanwhile, hospital admission flagged by a broader set of ICD codes with lower specificity could be subject to confirmatory medical record reviews, including causality assessment and root cause analysis, thereby potentially improving future patient safety. Clinically validated ICD code sets can serve as a proxy for ADE burden and guide the development of preventive strategies. ³⁰

4.5. Potential ways to further increase sensitivity of ADE detection

Ensuring an active pharmacovigilance service and employing varied approaches are critical for effectively detecting ADEs, with ICD codes playing a supplementary role in identifying potential ADEs. ³¹ To increase the detection of ADEs, ICD codes could be combined with other triggers, e.g., with electronic prescribing that indicate drug discontinuations or dose changes. ²² , ³² A Canadian study ²² found that combining an expanded ICD diagnostic code set with electronic prescribing data further increased the sensitivity of ADE detection. Moreover, advancements in machine learning and medical informatics might enhance the utility of ICD codes for future ADR detection. ¹³ , ³³

Furthermore, beyond using ICD codes to identify ADE‐related hospital admissions, these codes could also be applied earlier during emergency department visits, as suggested by Ruiz‐Ramos. ³⁴

4.6. Strengths and limitations of the study

The strengths of this study lie in its comprehensive examination of both hospital admissions and readmissions using ICD codes and clinical adjudication of ADEs. Unlike many studies that review medical records solely for (re)admissions flagged by ADE‐related ICD codes, which limits the ability to evaluate the sensitivity of these codes, our study estimated both the sensitivity and specificity of the ICD code sets developed by Hohl et al. across different settings. Additionally, we compared the performance of the ICD code sets in identifying preventable vs. nonpreventable ADEs, an analysis that has not been previously performed.

The main limitation lies in the small number of identified ADEs. Consequently, we were unable to assess the accuracy of individual ICD codes. Identifying a higher number of ADEs requires a very large sample size, which is uncommon in studies involving clinical adjudication using medical record reviews. An individual patient data meta‐analysis could address this challenge by combining datasets from studies that have identified ADEs leading to hospital admissions and recorded diagnoses of these admissions. Such an analysis would provide reliable information on the validity of various ICD codes and code sets for detecting ADEs and could highlight ICD codes that are indirectly linked to ADEs, but not yet included in the broad ICD code set.

In addition, the sample size was not determined through a formal calculation, as the study used datasets from previously published studies.

Another limitation stems from the variability in diagnoses recorded at different stages of patient care. Diagnoses documented at hospital admission frequently differ from the final diagnoses determined at discharge. As a result, the specificity of the ICD code set may vary depending on whether it is applied to admission or discharge diagnoses.

5. CONCLUSION

The study highlights that while a narrow ICD code set demonstrates high specificity, its sensitivity is limited. Broadening the ICD code set resulted in an 8–9‐fold increase in sensitivity, allowing for the identification of more ADE‐related admissions without significantly compromising specificity. Therefore, relying solely on a narrow ICD code set is insufficient; instead, incorporating multiple approaches, including the use of a broader ICD code set, is necessary for the identification of ADE‐related hospital admissions.

AUTHOR CONTRIBUTIONS

Daniala Weir and Fatma Karapinar‐Çarkit contributed to the study conception and design. Data analysis was performed by Zuzana Juhásová. All authors were involved in the interpretation of the results. The first draft of the manuscript was written by Zuzana Juhásová and all coauthors contributed to and approved the final manuscript.

CONFLICT OF INTEREST STATEMENT

The authors have no conflict of interests that are directly relevant to the content of the article.

Supporting information

TABLE S1 List of true positive ADE‐related ICD diagnoses (dataset from the Netherlands).

TABLE S2 List of true positive ADE‐related ICD diagnoses (dataset from the Czech Republic).

TABLE S3 Sensitivity and specificity of different code sets grouped by preventability of ADE‐related admissions

Juhásová Z, Karapinar‐Çarkit F, Weir DL. The use of international classification of diseases codes to identify hospital admissions linked with adverse drug events: Validation study. Br J Clin Pharmacol. 2025;91(10):2910‐2918. doi: 10.1002/bcp.70116

DATA AVAILABILITY STATEMENT

The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.