Class-level drug safety surveillance in hospital practice: A pharmacovigilance analysis of 84 therapeutic subclasses from a South Asian tertiary care center

Abstract

Objective

Post-marketing drug surveillance is critical for strengthening pharmacovigilance in low- and middle-income settings. This study provides the first class-level pharmacovigilance analysis from a South Asian tertiary hospital, generating safety signals across 84 drug subclasses and evaluating concordance between Proportional Reporting Ratio (PRR) and Reporting Odds Ratio (ROR) methods.

Methods

We conducted a retrospective cross-sectional analysis of electronic health records (2018–2022), comprising 718,088 drug administration events. Adverse drug reactions (ADRs) were identified using ICD-9-CM/ICD-10-CM codes. Disproportionality analyses using PRR and ROR were applied to each subclass.

Results

PRR identified positive signals in 5 subclasses, while ROR identified 7. The seven high-risk subclasses included antitubercular agents (PRR 3.24; ROR 4.14), anesthetic muscle relaxants (2.44; 2.84), aminoglycosides (2.24; 2.55), immunosuppressants (2.19; 2.47), neurologic sedatives (2.06; 2.29), antineoplastic cytotoxic agents (1.95; 2.15), and general anesthetic agents (1.88; 2.06). Concordance between PRR and ROR was high (96.4% agreement; κ ≈ 0.80; McNemar p = 0.48). Signal strength correlated inversely with overall drug exposure (ρ = −0.34).

Conclusion

This analysis establishes baseline class-level risk estimates in a developing pharmacovigilance environment, identifies disproportionately high-risk drug classes, and supports the robustness of PRR and ROR as complementary screening tools for regional drug safety monitoring.

Highlights

• •First South Asian class-level drug safety surveillance analysis.
• •Seven high-alert therapeutic classes identified from 718,088 cases.
• •PRR and ROR methods showed 96% agreement in signal detection.
• •High-volume drugs showed inverse associations masking potential risks.

1. Introduction

Pharmacovigilance comprises the detection, assessment, understanding, and prevention of adverse effects or any other drug-related problem.¹ Adverse drug reactions (ADRs) represent a substantial burden on healthcare systems, accounting for a significant percentage of hospital admissions and remaining a major contributor to morbidity and mortality.² Pre-market clinical trials are often constrained by size, duration, and patient diversity, making effective post-marketing surveillance essential.3, 4

Spontaneous reporting systems (SRS) allow healthcare professionals to report suspected ADRs to regulatory agencies.⁵ National and international databases, such as the US Food and Drug Administration's (FDA) Adverse Event Reporting System (FAERS) and the World Health Organization's (WHO) VigiBase, are real-world data repositories with millions of individual case safety reports.6., 7. However, it is estimated that as few as 1–10% of all ADRs are ever reported, and collected data is susceptible to reporting bias and confounding by indication.⁸ This necessitates statistical methods, such as signal detection algorithms (SDAs), to identify drug-event combinations that are disproportionately reported and warrant further investigation.⁹ The Proportional Reporting Ratio (PRR) and the Reporting Odds Ratio (ROR), two such SDAs, compare reporting rates for specific drug-event pairs against background rates.9, 10. Values exceeding thresholds (e.g., PRR ≥ 2) suggest safety signals.

This gap is critical in South Asia, where pharmacovigilance systems are developing and ADR reporting remains below 1%, compared to 5% globally.11, 12 The lack of systematic data from the developing world leads to a reliance on global databases, which may overlook or dilute signals driven by unique regional factors, such as disease burden, generic drug utilization, genetic diversity in metabolism, and resource constraints.4, 11 As such, relying solely on global pharmacovigilance data risks inappropriate signal prioritization and resource allocation.

In the present study, we analyzed 718,088 drug administrations (2018–2022) from a South Asian tertiary hospital, generating safety signals across 84 drug subclasses using PRR and ROR methods. This represents the first comprehensive, class-level pharmacovigilance analysis from this region.

2. Methods

This retrospective cross-sectional study used electronic health data from the Department of Internal Medicine, Aga Khan University Hospital (AKUH), Karachi, Pakistan, during 2018–2022. AKUH is a major 700-bed private, non-profit, teaching tertiary care facility serving a diverse and high-acuity patient population across Pakistan and the broader South Asian region. The study was approved by Aga Khan University's Ethics Review Committee under ERC# 2025–9035-35,501. Electronic data was solicited and extracted from two main sources: the Department of Pharmacy database and the Hospital Information Management System.

Adverse drug reaction (ADR) cases were identified using diagnostic codes from the International Classification of Diseases, 9th Revision, Clinical Modification (ICD-9-CM) (995.2, E930–E949, E933.1) and the 10th Revision (ICD-10-CM) (T36.0 × 5–T50.9 × 5, T45.1 × 5, T50.Z95, T78.2 × 5, T88.6, T88.7, Y40–Y59). We excluded non-pharmacologic agents (such as benign topical agents or infusions, alternative medicines, etc.) via a systematic review by two independent reviewers, with discrepancies resolved through discussion to reach consensus. Pharmaceutical agents were systematically classified across 84 major therapeutic subclasses, falling under the following eight primary categories: antimicrobials, cardiovascular agents, analgesics, anesthetics, neurologic agents, gastrointestinal medications, endocrine drugs, and other specialized agents (e.g., antineoplastics, biologicals, and antidotes).

For exposure calculation, each unique drug was counted once per admission regardless of frequency of administration (binary exposure: administered vs. not administered). When an ADR occurred during an admission, it was attributed to all drugs administered during that admission period. The total number of exposures was arbitrarily sub-classified as high exposure if greater than 10,000 exposures. Subclasses with ≤10,000 drug exposures were collectively categorized as low-to-moderate exposure, as the analysis focused on contrasting high-volume versus lower-volume prescribing patterns rather than granular exposure strata.

For signal detection, we used two disproportionality analysis methods: Proportional Reporting Ratio (PRR) and Reporting Odds Ratio (ROR).¹² A drug class-ADR combination was considered a positive signal if PRR ≥ 2.0, chi-square ≥ 3.84 (p < 0.05), and the number of cases ≥3. Signal criteria for ROR included a lower bound of 95% confidence interval > 1.0 and a number of cases ≥3.

2.1. Statistical analysis

Statistical analysis included basic descriptive statistics for all drug subclasses, with total drug exposures, ADR cases, and ADR proportions calculated. Categorical variables were expressed as frequencies and percentages, and continuous variables were expressed using medians and interquartile ranges. Agreement between PRR and ROR methods was assessed using the overall concordance rate (percentage agreement), Cohen's kappa (κ) coefficient for inter-rater reliability, and McNemar's test for marginal homogeneity of paired categorical data. The relationship between drug exposure (total use cases) and signal strength was evaluated using Spearman's rank correlation coefficient (ρ). Supplementary analyses included additional exploration of heterogeneity within the subclasses. All processing of the data and subsequent statistical analyses were performed using R statistical software (version 4.3.0) and RStudio integrated development environment.

3. Results

The analysis included 718,088 overall cases of drug administration across 84 unique subclasses, covering 528 unique drug preparations. For signal detection, we utilized both the Proportional Reporting Ratio (PRR) and the Reporting Odds Ratio (ROR) methods. Among the 84 subclasses evaluated, 5 (6.0%) returned a valid positive safety signal through PRR criteria and 7 (8.3%) through ROR criteria. Overall, the agreement between PRR and ROR methods was high (96.4% concordance, Cohen's κ ≈ 0.80, McNemar's p-value =0.48).

The full list of all the drug subclasses with generated signals is available as Appendix 1. Therapeutic subclasses that returned positive signals are reported in Table 1.

Table 1.

Drug subclasses with positive safety signals.

Therapeutic Class	Subclass	Drug ADR Cases	Drug Non-ADR Cases	Drug ADR Proportion (%)	PRR (95% CI)	ROR (95% CI)
Antimicrobial	Antitubercular	231	573	28.73	3.238 (2.904–3.612)	4.141 (3.553–4.825)
Anesthetic	Muscle Relaxant	311	1126	21.64	2.44 (2.211–2.693)	2.838 (2.502–3.219)
Antimicrobial	Aminoglycoside	194	780	19.92	2.243 (1.977–2.545)	2.552 (2.18–2.988)
Immunologic	Immunosuppressant	81	336	19.42	2.185 (1.797–2.658)	2.471 (1.939–3.15)
Neurologic	Sedative	179	801	18.27	2.056 (1.801–2.348)	2.293 (1.949–2.696)
Antineoplastic	Cytotoxic⁎	43	205	17.34	1.95 (1.486–2.559)	2.149 (1.547–2.986)
Anesthetic	Anesthetic Agent⁎	218	1088	16.69	1.88 (1.665–2.122)	2.056 (1.777–2.378)

^⁎Denotes a subclass with a positive ROR signal but not a positive PRR signal (i.e., PRR < 2.0).

As detailed in Table 1, the valid positive signals were highly concentrated in well-established, high-risk therapeutic subclasses. Antitubercular agents generated the highest disproportionality scores (PRR = 3.24, ROR = 4.14). Strong signals were also returned by anesthetic muscle relaxants (PRR = 2.44, ROR = 2.84) and aminoglycosides (PRR = 2.24, ROR = 2.55).

The 7 positive signals occurred in low drug exposure subclasses while high exposure subclasses showed no signals or inverse associations (PRR < 1.0). Commonly prescribed subclasses exhibited significant inverse associations (PRR < 1.0, p < 0.05), detailed in Table 2.

Table 2.

Drug subclasses with significant inverse associations (PRR < 1.0, p < 0.05), top 10 by total use cases.

Therapeutic Class	Subclass	Total Use Cases	PRR (95% CI)
Gastrointestinal	PPI	61,679	0.79 (0.767–0.814)
Analgesic	Non-Opioid	33,046	0.811 (0.78–0.844)
Antimicrobial	Cephalosporin	28,560	0.646 (0.616–0.677)
Analgesic	NSAID	20,369	0.839 (0.799–0.881)
Endocrine	Corticosteroid	16,980	0.933 (0.889–0.978)
Gastrointestinal	GI Agent	16,948	0.781 (0.74–0.825)
Endocrine	Insulin	15,909	0.908 (0.871–0.946)
Cardiovascular	CCB	13,382	0.909 (0.861–0.959)
Antimicrobial	Macrolide	12,135	0.703 (0.656–0.754)
Metabolic	Statin	10,757	0.786 (0.742–0.834)

In addition, there was a strong inverse correlation between signal strength and frequency of drug exposure (Spearman's ρ = −0.34, p < 0.001, for both PRR and ROR). Fig. 1 further shows the distribution of signals against their usage frequency.

Fig. 1.

Disproportionality metrics (PRR and ROR) versus total use cases.

Lastly, an exploration of heterogeneity between drug-level signals within our subclasses and contextualization of resultant aggregation effects is provided in supplementary materials.

4. Discussion

This is the first comprehensive class-level pharmacovigilance analysis from a South Asian hospital. We identified high-risk therapeutic categories, providing an important perspective on local drug safety patterns and allowing for comparison with established global pharmacovigilance.

We found the strongest disproportionality signal for antitubercular agents (ROR 4.14, PRR 3.24), perhaps driven by the mandatory, prolonged treatment using multidrug regimens (often 4–5 drugs¹³). We also found strong signals for anesthetic muscle relaxants (ROR 2.84, PRR 2.44), which are linked to specific, serious adverse events like anaphylaxis and malignant hyperthermia¹⁴ and aminoglycosides (ROR 2.55, PRR 2.24), which have known potential for nephrotoxicity and ototoxicity.¹⁵ This highlights the need for therapeutic drug monitoring and may also reflect a prescriber's reliance on older, more toxic agents due to limited availability or cost of newer alternatives. Immunosuppressants (ROR 2.47, PRR 2.19) showed expected risks with the complexity of managing transplant and immunocompromised patients. Their usage is linked to infection, malignancy, and organ-specific toxicity.¹⁶ The signal for sedatives (ROR 2.29, PRR 2.06) may reflect a range of adverse effects, from prolonged sedation or minor dizziness to serious outcomes such as respiratory depression, delirium, or falls.¹⁷ Lastly, the positive signals for antineoplastic cytotoxic agents (ROR 2.15) and general anesthetic agents (ROR 2.06) shows their inherently narrow therapeutic indices, and mechanism-based toxicities such as myelosuppression, mucositis, organ-specific damage, and peripheral neuropathy.18, 19 Illustrative drug-level findings are provided in the supplementary appendix.

Several of the most widely used drug classes showed statistically significant inverse associations (PRR < 1.0). This association for high-use agents like PPIs and non-opioid analgesics is likely influenced by clinical practice and reporting culture. Effective clinical management and prescriber familiarity may also lead to fewer ADRs in general. It is more plausible for this to be an artifact due to reporting practices and data structure as opposed to a truly superior safety profile. Many minor or common adverse events (e.g., headache, mild GI upset) associated with these drugs may be managed symptomatically by patients or clinicians without formal reporting or even documentation in the medical record as an ADR.2, 11 Furthermore, with co-prescription in a hospital setting, physicians may accredit an acute event to notoriously higher-risk agents than chronic-use or background drugs. Even when adverse events are documented, such attribution bias favors infamous drugs over common ones in polypharmacy.²⁰

Beyond these explanations, these patterns likely also reflect a ‘dilution effect’, whereby high drug-exposure denominators suppress apparent signals due to the sheer volume of prescriptions, sometimes also creating false ‘protective’ effects, an explanation further supported by the inverse correlation between drug administration frequency and signal strength (ρ = −0.34). This masking of potential harm in the most commonly used medicines can thus represent a significant public health challenge, as even a rare ADR can affect a large number of people due to the sheer volume of prescriptions. As such, it would be advisable to maintain consistency with best pharmacovigilance practices and not interpret PRR < 1.0 as true clinical safety.

Our analysis confirms a substantial degree of agreement between the PRR and ROR methods (96.4% concordance, Cohen's κ ≈ 0.80 suggesting that either test may be used in drug-safety surveillance as they are unlikely to identify differing primary safety signals, consistent with prior literature.⁹ Moreover, while the ROR algorithm showed slightly greater sensitivity, identifying two additional signals (cytotoxic agents and general anesthetic agents) at the PRR < 2.0 threshold, McNemar's test returned a p-value of 0.48, confirming a lack of systematic bias, and that neither test was significantly more liberal or conservative than the other in flagging safety signals.

Pakistan's pharmacovigilance system is developing. Our hospital-based data, derived from electronic records, likely represents a higher fidelity of reporting compared to the voluntary, passive reporting in many lower-resource settings within the country. Nationally, ADR reporting remains suboptimal due to limited awareness among healthcare professionals, low clinician-to-patient ratios, heavy workload, and a lack of dedicated clinical pharmacy services.11, 21 Still, our positive signal findings provide quantitative evidence that retrospective EHR screening is feasible and can inform risk minimization strategies in developing pharmacovigilance systems.

Our findings have important implications for regulatory agencies and hospital pharmacovigilance teams. We identify high-risk therapeutic categories (e.g., anti-TB agents, anesthetic muscle relaxants, aminoglycosides) warranting heightened, targeted monitoring and resource prioritization within our setting. The presence of inverse associations for high-use drugs shows that disproportionality metrics alone might be insufficient for comprehensive risk assessment. Effective surveillance might need drug exposure data, absolute ADR counts, and clinical context alongside these metrics. Critically, these metrics should not be considered definitive evidence of safety. Future research should build on this analysis by investigating individual drugs.

5. Limitations

This study has several limitations inherent to the use of spontaneous reporting data. First, spontaneous reporting data cannot calculate true incidence or risk due to underreporting. Based on data from a large, private tertiary care center, which may not fully reflect safety patterns in public or community settings in Pakistan, our results might not entirely generalize. In Pakistan, private tertiary centers tend to cater to patients of a higher socioeconomic class, with potentially different susceptibility profiles, than the vast majority of the population belongs to. Thus patterns of drug accessibility and administration, along with ADR reporting are likely to be more robust than the public and rural settings which face higher patient volumes and added capital and resource restrictions, but which may also result in different prescription patterns.

Lastly, using classes and subclasses for our analysis might incur the ecological fallacy; one drug may drive the class-level signal, or a drug-level signal might be muted by an otherwise safe-appearing class. This assumption is further explored in supplementary materials. Additionally, our model attributed adverse events to all drugs that were given during an index admission. Although this method aligns with surveillance practices, it can lead to common attribution among frequently co-prescribed drugs. This may also explain the observed inverse correlation between drug exposure and signal strength. We also lacked patient-level demographic and clinical characteristics (age, sex, comorbidities, renal/hepatic function), which are important effect modifiers for ADR risk. Our dataset also lacked the clinical detail needed to assess ADR severity, preventability, or the influence of confounding by indication. Also, while our exclusion of non-pharmacologic agents was conducted by two independent reviewers, it lacked formal inter-rater reliability testing.

6. Conclusion

This class-level analysis from a South Asian tertiary hospital identified high-risk drug categories and regional safety patterns. The high concordance between PRR and ROR (κ ≈ 0.80), validates their use as an initial screening tool in resource-constrained settings. Effective local pharmacovigilance requires a shift to active surveillance strategies incorporating monitoring of high-risk classes and the usage of local drug exposure data, absolute ADR counts, and clinical context for accurate signal interpretation.

Future studies should focus on incorporating drug-level granularity, with particular focus on high-risk drugs identified in this study (antitubercular and aminoglycoside agents) to add to literature on the safety profile of medicines of essential use in the population. Additionally, a focus on linking spontaneous reporting with laboratory findings of individuals experiencing ADRs can help accurately identify offending agents and validate safety signals. Finally, integration of validated hospital-level data with national pharmacovigilance databases can facilitate mapping these findings and determining whether they persist in the broader population -and help inform further national pharmaceutical guidelines and safety alerts.

CRediT authorship contribution statement

Abdullah Umer: Writing – original draft, Visualization, Formal analysis, Conceptualization. Adam Khan Mohmand: Writing – original draft. Naimal Mouzam Khan: Writing – original draft. Ainan Arshad: Writing – review & editing, Validation, Supervision. Sher M. Sethi: Writing – review & editing, Validation, Supervision.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

Complete list of Sub-class level signals.

Drug-Level Disproportionality Analysis