Digital Medicines Information at National Scale: Search Behaviour and System Performance of MediVerify Across 1.5 Million Queries

1 Introduction

• MediVerify was developed to address scalability, update-frequency, and searchability limits in Sri Lanka’s publicly accessible PDF medicines registry.The platform provides a real-time searchable registry of NMRA-approved products.Source: §1 · p. 2
• Long-term national medicine-search patterns had not been characterised, although a full year of queries could reveal information needs and policy alignment.The question includes potential mismatches between public queries and essential-medicine priorities.Source: §1 · p. 2
• Published analyses of comparable national medicines platforms in LMICs remain sparse, despite related work in England, the U.S., and Singapore.Prior research also described unmet demand for medicines information and frequent absence of highly prescribed medicines from essential medicines lists.Source: §1 · p. 2
• The study presents a year-long analysis of MediVerify and aims to characterise search behaviour, assess essential-medicines alignment, and evaluate platform performance.The authors describe it as the first large-scale analysis of national medicine search behaviour in an LMIC, to their knowledge.Source: §1 · p. 2

2.1 Research design

• The non-experimental retrospective observational design analysed pre-existing system data without intervention or manipulation of study variables.Queries were collected from 11 July 2024 through 10 July 2025; 1,497,304 anonymised entries were mapped against 11,933 NMRA product entries.Source: §2.1 · p. 2, §2.2 · p. 2

2.2 Study design and data source

• The dataset comprised all 1,497,304 anonymised searches submitted during the one-year study period, with the NMRA’s 11,933-product registry as the reference database.No user-identifying data were collected.Source: §2.2 · p. 2

2.3 Query processing and classification

• Queries were normalised, matched against generic and registered brand names, and assigned using exact or Levenshtein fuzzy matching at a maximum edit distance of five.The analysis also classified therapeutic categories, identified zero-query medicines, manually reviewed ambiguous mappings, and cross-referenced leading medicines against WHO and Sri Lankan essential medicines lists.Source: §2.3 · pp. 2–3

2.4 System performance analysis

• System performance analysis summarised query latency and examined its relationship with edit distance, while energy use was estimated from processing time and assumed server power.Energy was extrapolated to kWh per million searches; carbon estimates used Sri Lanka’s grid-emission factor, and CPU utilisation was analysed across months and study quartiles with non-parametric tests.Source: §2.4 · p. 3

3.1.1 Distribution of searches by therapeutic category

• Vitamins/minerals accounted for 11.74% of searches, antibiotics 10.57%, diabetes-related drugs 7.02%, and antihypertensives 6.84%.Together, these four categories represented more than 36% of all searches.Source: §3.1.1 · p. 3, Figure 1 · p. 4

3.1.2 Most searched individual medicines

• The top 20 medicines accounted for approximately 37.3% of queries; rosuvastatin led with 48,812 searches (approximately 3.26%).Paracetamol ranked second with 45,967 searches (3.07%), followed by glimepiride with 45,668 (3.05%).Source: §3.1.2 · p. 3

3.1.3 Zero-query registered medicines

• Approximately 40%–50% of registered medicines were never queried during the study year, largely specialised or seldom-used products.Examples included oncology drugs, biologic therapies, antidotes, paediatric formulations, and hospital-restricted medicines.Source: §3.1.3 · p. 3

3.1.4 High-frequency zero-result searches

• Approximately 1% of searches returned no results, including unregistered products, possible registry or naming gaps, brand–generic mismatches, spelling variants, and non-medicinal terms.Mebendazole was identified as apparently absent from the NMRA registry during the study period, suggesting a regulatory gap or naming mismatch.Source: §3.1.4 · p. 3

3.1.5 Interquartile variability of top 20 searched medicines

• Search volumes in Q3 and Q4 were significantly lower than in Q1 and Q2 (p < 0.001); Q4 was also lower than Q3 (p = 0.0047).The overall difference across quarters was significant (Friedman χ2 = 43.32, p = 2.10 × 10^-9); Q1 and Q2 did not differ significantly (adjusted p = 0.091).Source: §3.1.5 · pp. 3–4, Figure 2 · p. 5

3.2.1 Essential medicines list cross-reference

• Nine of the top 20 queried medicines were absent from both essential medicines lists, and nearly 70% were absent from one or both.Rosuvastatin and glimepiride were among the nine absent from both the 2025 WHO Model EML and the 2022 Sri Lanka National EML.Source: §3.2.1 · p. 5

3.3.1 Search accuracy and fuzzy matching performance

• Approximately 85% of searches matched exactly, while about 2% had edit distance ≥3; edit distance and latency were positively correlated (r ≈ 0.69, p < 0.001).The higher-distance queries were relatively rare.Source: §3.3.1 · p. 5

3.3.2 System latency

• Median latency was approximately 8 ms, while the 99th percentile was approximately 800 ms and fewer than 1% of queries exceeded 1 s.Latency was right-skewed, and slower outliers were generally associated with higher edit distances or many matching results.Source: §3.3.2 · p. 5

3.3.3 Computational and energy footprint

• Approximately 0.12 kWh per million queries was the estimated energy consumption; mean CPU utilisation was 0.358% over the year.Energy estimates used simplified power assumptions and were described as order-of-magnitude approximations.Source: §3.3.3 · pp. 5–6

• Interest in chronic-disease medicines, antibiotics, and supplements broadly aligned with patterns at other medicines-information platforms; cited comparisons include NHS Open Prescribing [3] and Moon et al. [7].The discussion notes that frequently searched rosuvastatin, glimepiride, and amlodipine are consistent with frequently prescribed drug classes across LMICs, regardless of EML status [7].Source: §4 · p. 6
• Approximately 85% of searches were exact matches, and over 90% of queries completed in under 50 ms; higher edit distances were associated with longer response times.Queries with edit distance ≥3 accounted for the largest latencies, up to 1–3 s, but were uncommon; overall query success was approximately 99%.Source: Figure 3 · p. 6, Figure 4 · p. 7, Figure 5 · p. 7
• In sub-Saharan African LMIC settings, substantial public demand exists for medicines outside formal essential medicines frameworks; search analytics may provide a near-real-time proxy for evolving public demand and prescribing (8, 15).This could complement episodic expert-consensus processes currently underpinning EML updates.Source: §4 · p. 7
• Zero-result queries made up approximately 1% of searches, while 40%–50% of registered medicines were never queried, a long-tail pattern also reported for NHS Open Prescribing.The discussion suggests monitoring zero-result queries to guide database improvements such as adding synonyms and spelling variants.Source: §4 · p. 8
• Estimated energy use was approximately 0.12 kWh per million queries, which the discussion presents as evidence of a very low environmental footprint for national-scale search.The energy estimate relies on simplified assumptions.Source: §4 · p. 8, Figure 6 · p. 8

5.1 National medicine information-seeking behaviour

• Searches concentrated in a few therapeutic categories, varied strongly across quarters, and zero-result monitoring can identify registry gaps for database updates or public education.The authors also identify the data as actionable for platform maintenance and scheduling public-health communication.Source: §5.1 · p. 8

5.2 Essential medicines policy alignment

• Approximately 70% of the top 20 searched medicines were absent from one or both essential medicines lists.The conclusions call for demand data to inform periodic EML revision and integration into the review cycle.Source: §5.2 · pp. 8–9

5.3 System performance and sustainability

• >1.49 million year-one queries were handled with 8 ms median latency, >99% query resolution, and approximately 0.12 kWh per million queries.The authors describe this combination as high utilisation with minimal computational and environmental overhead.Source: §5.3 · p. 9

5.4.1 National medicines regulatory authority (NMRA)

• The NMRA recommendations include adding synonym and brand variants, reviewing zero-result logs monthly, addressing registry gaps, and expanding medicine information.The proposed expansion includes patient information leaflets and regulatory approval-status history.Source: §5.4.1 · p. 9

5.4.2 Ministry of health and nutrition

• The Ministry recommendations are to review the national EML using query demand, integrate analytics into the Pharmaceutical Management Information System, and share the model regionally.The proposed sharing is with the WHO SEARO network to facilitate adoption by other LMIC regulatory authorities.Source: §5.4.2 · p. 9

5.4.3 Provincial and sub-national health directorates

• Provincial directorates are urged to use targeted campaigns, provincial planning analytics, and pharmacist and community health-worker training on MediVerify.The recommendations include raising awareness of essential medicines in the zero-query category and using MediVerify as a first-line information resource.Source: §5.4.3 · p. 9

6 Limitations

• Search queries may not directly reflect actual medicine utilisation or prescribing patterns.The authors caution against treating search behaviour as a direct measure of use or prescribing.Source: §6 · p. 9
• Without user-level demographic data, the study cannot interpret patterns across population subgroups.Source: §6 · p. 9
• Registry completeness and naming conventions may affect zero-result and zero-query findings, while energy estimates are order-of-magnitude approximations.The energy estimates rely on simplified assumptions rather than precise measurements.Source: §6 · p. 9
• Findings come from one national platform in one LMIC context, so generalisability to other regulatory settings requires validation.Source: §6 · p. 9

Ethics statement

• Ethical approval and written informed consent were not required under local legislation and institutional requirements.Source: Ethics statement · p. 9

Generative AI statement

• The authors used ChatGPT version 5 for image generation and linguistic review, and Frontiers generated figure alt text with AI support.Source: Generative AI statement · p. 10

• The article’s claims are solely the authors’ and do not necessarily represent their organizations, the publisher, editors, or reviewers.Source: Publisher’s note · p. 10
• The publisher does not guarantee or endorse any evaluated product or claim made by its manufacturer.Source: Publisher’s note · p. 10