ICD-10 Codes to Identify Adverse Drug Events Associated with Antibiotics in Administrative Data

TL;DR

This paper describes a method to identify adverse drug events linked to antibiotics using hospital diagnostic codes, helping track antibiotic-related harms in healthcare data.

Contribution

A novel methodology for developing and validating ICD-10 codes to identify antibiotic-associated adverse drug events in administrative data.

Findings

72 ICD-10 codes were classified as likely antibiotic-associated.

Inter-rater reliability was assessed using Kappa scores with 95% confidence intervals.

The code list can improve capturing antibiotic-related adverse events in administrative data.

Abstract

Antibiotics are among the most used therapeutics in primary care, and while their benefits are clear, the potential harms related to adverse drug events (ADEs) cannot be ignored. We outline the creation of a comprehensive list of diagnostic codes describing antibiotic-associated ADEs resulting in presentations to acute care hospitals. Methods: Previously published ADE codes were used to link BC hospitalizations to prior outpatient antibiotic prescriptions and were restricted based on whether patients received an antibiotic within a month prior to the ADE-related hospitalization. The code list was reviewed by two clinical experts independently for the likelihood of being antibiotic-associated. The inter-rater reliability was calculated using Kappa scores with 95% confidence intervals (CIs). Results: Of the 695 ICD-10 ADE codes with evidence of recent antibiotic administration, 72, 68, and 555 codes were considered likely, possibly, and unlikely antibiotic-associated, respectively. Conclusions: We outline a methodology for developing an ICD-10 code list for antibiotic-associated ADEs severe enough to warrant hospital admission. This will help to improve the use of administrative data to capture antibiotic-associated ADEs.

Full text

1. Introduction

Outpatient antibiotic use is common, but may confer health risks that compromise its therapeutic benefits, including allergic reactions (i.e., anaphylaxis and skin rash), disruption of the gut microbiome (i.e., Clostridioides difficile infection) and drug interactions [1,2,3]. Adverse drug events (ADEs) associated with outpatient antimicrobial medications are responsible for more than 280,000 patient visits to emergency departments and 26,000 admissions across Canada per year, resulting in over USD $400 M in annual healthcare expenditures [4,5,6,7,8]. These costs do not take into account the additional burden of antibiotic-resistant infections and attributable deaths. The Council of Canadian Academies reported that in 2018, 26% of the approximately 1 million bacterial infections in Canada were resistant to first-line treatments, with resistant infections being responsible for over 14,000 deaths, 5400 of which were directly attributable to antimicrobial resistance [9]. Capturing data on antibiotic-related adverse drug events observed in clinical practice is important to understand the real-world clinical benefits and harms of antibiotics, and inform drug regulation, prescribing guidelines, and efforts to curb clinical harm and antibiotic resistance.

Administrative databases are readily available, inexpensive and can provide population-level health data on important health outcomes. In Canada, the Discharge Abstracts Database (DAD) captures hospital administrative data including up to 25 hospital diagnoses [10]. In the past, we produced a list of ICD-10 codes to identify ADE-related emergency department visits [11] and evaluated the code set for its completeness in identifying a broad range of ADEs [12]. In this work, we aim to identify a list of ICD-10 codes to identify ADEs specifically related to outpatient antibiotic dispensations that are severe enough to require hospitalization.

2. Materials and Methods

2.1. Identifying Potential Antibiotic-Associated Adverse Drug Events

We focused on adverse drug events (harm caused by appropriate or inappropriate use of a drug) as opposed to adverse drug reactions (harm caused by a drug under appropriate use) as empiric use of antibiotics is common practice and therefore assessing appropriateness is often not possible based on a lack of laboratory data. We identified potential antibiotic-associated ADEs by linking British Columbian administrative hospitalization (DAD) data with provincial medication dispensing data (PharmaNet) from 1 January 2001 to 31 December 2020 [13,14]. We started by identifying all hospital healthcare encounters coded with an ADE-related ICD-10 diagnostic code with the same first three digits as those identified by Hohl et al. in a systematic review of the literature [11,12,15].

We excluded ADE-related diagnostic codes describing events that occurred during a patient’s hospital stay (i.e., diagnosis type 2 codes—“post-admit comorbidity—conditions that arise following admission and satisfy the requirements for determining comorbidity”) [16]. We then excluded all healthcare encounters that were not preceded with one or more outpatient antibiotic dispensations within 30 days of the healthcare encounter. Analyses were performed using R Studio 2023.12.0.

This restricted code set was then assessed by two independent reviewers for the likelihood of the ADE being antibiotic-associated. A physician with expertise in infectious diseases and microbiology (JG) and a clinical pharmacist (FM), both with extensive clinical experience practicing in British Columbia, independently rated the certainty by which any antibiotic could be responsible for the ICD-10 diagnosis coded in a subsequent hospital visit. Reviewers considered ADEs that could be directly (i.e., toxicity or drug interaction) or indirectly attributed to the use of antibiotics (i.e., dysbiosis or drug interaction). Reviewers used the Hohl et al. coding system (Table 1) [11]. Any discrepancy in the assigned causality rating for each ICD-10 code was discussed between the two reviewers, and a third clinical reviewer (CMH) where necessary, until consensus was reached.

2.2. Rating Inter-Rater Reliability

We calculated a raw and a weighted Cohen’s Kappa statistic to determine inter-rater reliability of the two reviewers’ initial ratings prior to reaching consensus. Causality codes were categorized into Likely (A, B, or C), Possible (D or E), and Unlikely (U or V). “Categorical agreement” was defined as agreement across the Likely, Possible, and Unlikely categories and “essential agreement” was defined as agreement across the A1, A2, B1, B2, C, D, E, U, and V categories. To calculate the weighted Kappa statistic, a linear weighting matrix was used to weigh differences between Likely and Unlikely more heavily than differences between Likely and Possible. Both raw and weighted Kappa statistics were calculated based on categorical agreement.

3. Results

Among the original list of 827 ICD-10 codes used in the literature for identifying ADEs, 695 different ADE-related codes were identified by linking provincial administrative hospital admission data with 30-day antibiotic dispensation data. There were 99 essential discrepancies (14.2%) and 97 categorical discrepancies (14.0%) between the two raters prior to reaching consensus. The unweighted categorical agreement between the two reviewers was “weak” (k = 0.54, 95% CI 0.45–0.62), but the weighted categorical agreement was “moderate” (k = 0.60, 95% CI 0.52–0.68) [17]. After facilitated discussion and an additional round of review, complete consensus on the code list was reached.

We identified 72 ICD-10 codes (10.4%) that “Likely” identified antibiotic-associated ADEs (categories A1, A2, and B1 and B2 and C) (Table 1). Another 68 ICD-10 codes (9.8%) that “Possibly” identified potential antibiotic-associated ADEs (categories E and D). The remaining 555 ICD-10 codes (79.9%) were deemed “Unlikely” to characterize an antibiotic-associated ADE (categories U and V). Some illustrative examples of “Unlikely” diagnostic categories included, but were not limited to, gastrointestinal ulcers (K25–K28), mental or behavioral disorders due to specified drugs (F11–F19), or external cause of morbidity codes specifying non-antibiotic drugs (Y41–Y88).

The “Likely” and “Possible” ICD-10 codes themselves can be found in Table 2 and Table 3, respectively.

4. Discussion

Our objective was to identify an ICD-10 code set among patients admitted to hospital for an ADE related to antibiotic use within the prior 30 days. Two expert reviewers adapted a previously published causal ranking system to identify the likelihood that an ADE was associated with antibiotic use. To our knowledge, this is the first study to develop an ICD-10 code set that can be applied to administrative healthcare records to identify the incidence of antibiotic-associated ADEs. These codes can be used to investigate the incidence of recognized antibiotic-associated ADEs and to better understand the antibiotic classes posing the greatest risk of subsequent antibiotic-associated ADE-related hospital admission.

In Canada, discharge diagnoses are coded retrospectively by administrative coding specialists from patient charts and the number of discharge diagnoses coded per visit varies by province [18,19]. The way these fields are coded varies by province and country such that our results may not be generalizable outside of British Columbia. Future studies should apply the antibiotic ADE diagnostic codes identified in this work to their own administrative data to refine and validate them in other health systems. Also, while administrative data generated by acute care encounters with the healthcare system have been impactful in understanding how medications are used in the real-world setting and in generating safety signals, few confirmed ADEs are documented within them. This was demonstrated by Wickham et al. when comparing ADEs identified in administrative data with confirmed ADEs identified prospectively by care teams in three prospective cohort studies in BC (although hospital data performed the best out of the administrative datasets investigated) [12]. Thus, while our intent was to identify antibiotic-related ADEs using ICD-10 codes, these codes can only identify ADEs caused by antibiotics that were recognized and documented by the care team [20]. Ideally, they should be validated in future studies against prospectively identified and confirmed antibiotic-related ADEs. Additionally, as the diagnostic code set comprises ICD-10 codes used in hospitals, they may not capture the ADEs that patients might present with at community-based clinics (where ICD-9 codes are used). However, the code set outlined in this paper will capture the most severe and costly ADEs that patients seek care for following antibiotic use.

Future research should validate this code set by applying it to identify antibiotic-associated ADEs in records, which is currently underway in BC. This paper provides important methodological guidance on how a code set may be generated, based on real-world coded data. Our work can also be used as a starting point to develop an explicit code set for community-based ADEs, and to identify and track hospital admissions for antibiotic-related ADEs. This will allow us to identify geographic variation and time trends as new antibiotics are licensed, and prescribing guidelines vary. While, undoubtedly, the proposed code set will require iteration, it provides a starting point for improved surveillance to identify, quantify, and compare antibiotic-associated harms.

5. Conclusions

We used community medication dispensing data linked with subsequent hospitalization data, along with clinical review and iterative consultation, to adapt a previously published list of antibiotic-associated ADE diagnostic codes. The list shared in this work may help identify antibiotic-associated adverse drug events in hospital inpatient records in administrative data, improving the ability to capture this important outcome when assessing the benefits and risks of antibiotic administration in patient populations.