Total Mendeley and Citeulike bookmarks.
Paper's citation count computed by Dimensions.
Sum of Facebook, Twitter, Reddit and Wikipedia activity.
Roles Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Supervision, Validation, Writing – original draft, Writing – review & editing
Affiliations Faculty of Medicine and Health Sciences, An-Najah National University, Nablus, Palestine, Department of Internal Medicine, Palestine Medical Complex, Ramallah, Palestine, Ministry of Health, Ramallah, Palestine
https://orcid.org/0009-0009-9721-6249
Roles Data curation, Validation, Writing – review & editing
Affiliation Faculty of Medicine and Health Sciences, An-Najah National University, Nablus, Palestine
Roles Data curation, Visualization, Writing – review & editing
Affiliation Ministry of Health, Ramallah, Palestine
Snakebite surveillance often relies on routine hospital records that were not designed to count events: they are duplicated, distributed across departments and repeated across hospitals, so the count depends on whether rows, encounters, patients or episodes are taken as the unit. We made each reconstruction step explicit.
We analysed all records carrying ICD-10 T63 snake-venom codes from 14 public hospitals in the West Bank, Palestine, between 14 March 2021 and 26 December 2025. The extract contained 362 rows. Removing rows identical on every field left 292; consolidating records sharing hospital, patient file number and visit date gave 265 registry encounters and 184 hospital-specific patient records. A deterministic rule fixed before the results — exact patient file number with concordant full name, recorded sex and recorded age — produced 172 linked patients through 12 cross-hospital links. A non-prespecified 14-day algorithm grouped these encounters into 178 algorithmically reconstructed episodes ; counts were 183 and 177 under 7- and 30-day windows. A 13th potential link found after the analysis lock is reported only as a sensitivity analysis (171 patients); across 11 linkage specifications tested post hoc during revision, the count was 171 or 172. A nested chart review at three hospitals retrieved 35 of 37 eligible charts; 31 patients were clinically assessable on arrival. Documented systemic manifestations appeared in 12, antivenom was given to 17 and intensive care to 11; there was no in-hospital death, and four documented deaths preceded arrival.
Rows, encounters, patients and reconstructed episodes are different quantities, and the same extract yields substantially different counts at each step. Routine records can support surveillance reconstruction when every step is explicit, but these are counts of hospital-recorded care, not estimates of snakebite incidence or burden, and episode counts are algorithmic reconstructions rather than clinically validated bite counts.
Snakebite is counted, in most countries, from records that hospitals keep for administration rather than for surveillance. The same patient may generate several records: one for each department involved, another if the patient returns, and another at a second hospital. We took every record carrying a snake-venom diagnosis code from 14 public hospitals in the West Bank over almost five years and rebuilt the counts step by step. We removed rows that were identical on every field; combined records from the same hospital, for the same patient, on the same recorded date; linked patients across hospitals using a rule fixed in advance; and finally grouped visits close together in time into reconstructed episodes of care. Each step changed the total, so reading the same data as rows, encounters, patients or reconstructed episodes gives very different numbers. We also read the hospital charts of 35 patients at three hospitals to describe what clinicians recorded. Those charts cannot establish that each coded record was a snakebite, and we do not claim that they do. All of these numbers describe care recorded in hospitals, not how many snakebites occur in the population.
Citation: Taha HM, Taha KH, Abu Muhsen AH (2026) Snakebite records in routine hospital data: Deduplication, encounter consolidation, patient linkage and algorithmic episode reconstruction in 14 Public Hospitals, West Bank, Palestine, 2021–2025, with a nested chart review. PLoS Negl Trop Dis 20(11): e0014897. https://doi.org/10.1371/journal.pntd.0014897
Editor: Emeka John Dingwoke, UNESCO International Center for Biotechnology, NIGERIA
Received: August 1, 2026; Accepted: November 12, 2026; Published: November 19, 2026
Copyright: © 2026 Taha et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: The patient-level data underlying this study cannot be made publicly available because they contain sensitive clinical information derived from routinely collected healthcare records and are subject to institutional, ethical, confidentiality, privacy and data-governance restrictions. The derived analysis dataset is pseudonymised rather than anonymised: it carries no names, file numbers, dates of birth or exact dates, but retains pseudonymous patient and site identifiers and day offsets, and within a nested subset of 35 records drawn from three hospitals indirect identifiers could contribute to re-identification. Enquiries or requests concerning potential access to the restricted patient-level study data should be directed to the Institutional Research Ethics Committee (IREC) of Palestine Medical Complex (PMC), Palestinian Ministry of Health, Ramallah, Palestine, at pmc@moh.ps , quoting IREC reference PMCQu-22-04-2026. Submission of a request does not guarantee access. Each request is considered on a case-by-case basis through the applicable institutional process and remains subject to all applicable ethical, institutional, confidentiality, privacy, data-governance and authorisation requirements. No author is the sole individual responsible for ensuring access to these data. Analysis code and non-patient-level reproducibility materials are publicly available at https://doi.org/10.5281/zenodo.23002338 (version 1.0). Version 1.1 of the same record, at https://doi.org/10.5281/zenodo.24015979 , adds the code, a synthetic fixture and the aggregate output of the revision-stage linkage sensitivity analysis; running that code on study data requires the restricted identifier-bearing extract. Aggregate results sufficient to reproduce the reported summary findings are provided in the Supporting Information .
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
Snakebite envenoming is a WHO-listed neglected tropical disease and a time-critical medical emergency, responsible for an estimated 81,000–138,000 deaths and roughly three times as many permanent disabilities each year [ 1 , 2 ]. The WHO strategy for prevention and control sets a target of halving snakebite mortality and disability by 2030 and identifies weak surveillance as a principal obstacle: in most affected settings the true incidence, the distribution of care and the quality of clinical management are unknown [ 3 , 4 ]. The populations at greatest risk are typically those with the least access to appropriate care [ 5 , 6 ].
Snakebite surveillance is usually described as a problem of missing data. In many countries the more immediate problem is different: hospital records exist, but they cannot be interpreted as they stand. Routine hospital information systems are built for administration rather than epidemiology, and they share a set of failure modes that are both predictable and addressable. Cases are captured only when a patient attends, is documented and is coded in a retrievable way, so ascertainment is incomplete and selective by construction, and the resulting count is a measure conditioned on care-seeking, documentation and coding rather than a population incidence estimate. Where hospital and population sources have been compared directly, hospital records captured substantially fewer snakebites than community surveys [ 7 , 8 ], and routine reporting captured only a small fraction of expected snakebite deaths [ 9 ]. Coding conventions rarely distinguish a clinical episode from a hospital encounter, so continuing care and inter-hospital movement inflate counts unless they are explicitly reconciled. Patient identity may also be difficult to reconcile across facilities when identifier architecture is undocumented or when the same individual appears under different identifiers; linkage rules therefore need to be explicit and their uncertainty acknowledged [ 10 ]. Finally, the clinical detail needed to judge envenoming severity or the appropriateness of antivenom is generally absent from coded fields altogether. These processes can alter surveillance counts, although their direction and magnitude are not always identifiable from routine extracts.
Evidence from the eastern Mediterranean region is sparse. Daboia palaestinae, the regionally dominant medically important species, produces a predominantly haemotoxic syndrome with local tissue injury, and its clinical management has been described mainly in single-centre series from neighbouring health systems [ 11 , 12 ]. Published patient-level multicentre evidence on snakebite from the West Bank is limited. This study combines routine hospital records linked across facilities with case-level clinical review. That combination makes the two data sources complementary: registry data permit reconstruction of documented patients, hospitals and recorded dates, but cannot describe envenoming severity, treatment decisions or outcomes; chart review describes severity and treatment but cannot establish a denominator.
Routine hospital records are the default source for snakebite surveillance where no dedicated registry exists, yet the unit they produce is administrative. A record documents a departmental entry, not a patient and not a bite. Where records are duplicated, distributed across departments and repeated across hospitals, counts derived from them may describe the recording system as much as the clinical problem.
We therefore set out to make each counting step explicit and to report its effect: how many raw records remain after exact deduplication; how many belong to the same patient within and across hospitals; and how many groups of encounters an algorithmic episode rule produces. We did not set out to validate diagnostic coding, to estimate population incidence, or to establish that coded records represent envenoming, and the design does not support those aims.
The study was approved by the Institutional Research Ethics Committee of Palestine Medical Complex, Ramallah (reference PMCQu-22-04-2026; date of ethical approval 20 April 2026), which classified it as minimal-risk retrospective observational research using routinely collected records and waived the requirement for individual informed consent. No other ethical determination or amendment is relied upon.
Retrospective analysis of routinely collected hospital records from 14 public hospitals of the Palestinian Ministry of Health in the West Bank, covering 14 March 2021 to 26 December 2025, with a nested chart review at three of those hospitals. No participant was contacted and no clinical care was altered.
Records were obtained as a single hospital information system report extracted on 26 April 2026. The extract contained eight fields: patient file number, patient name, sex, age, department, ICD-10 code, code description and visit date. It contained no time-of-day variable, no outcome field, no bite date, and no field recording referral, transfer, readmission or follow-up status. The age field was calculated automatically by the hospital information system from each patient's date of birth at the extraction date; it is therefore age at extraction, not age at each encounter.
Records were ascertained by ICD-10 T63 snake-venom codes; eight distinct codes appeared ( S1 Table ). Two encounters carried species-denominated subcodes, which we treat as coding labels only: the code set cannot identify a species. Coding is the ascertainment criterion; it is not clinical confirmation of snakebite or of envenoming. No exclusion was applied by age, sex, department or outcome.
Rows identical on every extracted field were removed as exact duplicates. Remaining rows sharing hospital, patient file number and visit date were consolidated into one registry encounter , with contributing departments retained; multiple departmental rows could occur within one such combination. A patient–date key without hospital would merge records held by different hospitals for the same patient on the same date; five patient–date keys spanned two hospitals, so hospital forms part of the encounter key. Because no time-of-day variable or attendance identifier exists, a registry encounter is an algorithmically defined unit, and the data cannot establish that it corresponds to exactly one distinct clinical attendance. Every processing step, with the count before and after, is given in S2 Table .
Hospital-specific patient records were linked into globally linked patients by a deterministic rule fixed before the results were produced: an exact match on the patient file number together with concordance of the full name, the recorded sex and the recorded age. The rule is a conjunction; no single agreeing field could produce a link.
The normalisation actually applied to names during the original analysis is not documented, and the authors confirm, after reviewing all available study files, that no further contemporaneous record of the normalisation, of candidate-pair generation or of adjudication exists. The normalisation given in S3 Text is a revision-stage reconstruction that reproduces the locked result exactly, and is labelled as such rather than presented as the historical implementation; procedural detail that was not retained is not reconstructed here. A count of candidate pairs considered or rejected during the original linkage was not retained. The candidate-pair space analysed in the revision-stage sensitivity analysis (section 6) was reconstructed de novo during revision from the retained original extract; it is not a reconstruction of the original adjudication. All names in the extract are recorded in Arabic script, so variation between scripts is not observable in these data; this does not establish that such variation never arose before data entry.
The analytic extracts contained no identifier demonstrated to be valid across the hospital network. A national identification number exists in the source hospital information system but was not included in the extracts available for this analysis , so no independent reference standard was available against which to test linkage. In the study extract, the same patient file number persisted for the same patient across distinct hospitals; we describe this only as observed behaviour and make no claim about the underlying identifier architecture, which is not documented. False-negative linkage could not be quantified.
One further potential cross-hospital link was identified after the analysis lock: the same four-token full name, sex and recorded age, but different patient file numbers. It is strongly suggestive of one individual, falls outside the locked rule, and cannot be resolved without an authoritative identifier. Accepting it yields 171 patients, reported only as a labelled sensitivity analysis ( S4 Table ).
In response to peer review, we examined whether the number of linked patients depended on the choice of linkage rule. This analysis was not prespecified, is not part of the primary analysis and does not replace the locked result. Of the three approaches to evaluating linkage quality described by Harron and colleagues — comparison with a gold-standard subset, comparison of linked with unlinked records, and sensitivity analysis of results to changes in the linkage procedure [ 13 ] — the approach feasible with these data was sensitivity analysis to assess how strongly the linked-patient count depended on plausible changes in the linkage procedure; without a reference standard, this analysis cannot quantify false or missed links. The candidate-pair space was reconstructed de novo during revision from the retained original extract. A candidate pair was defined as an unordered pair of hospital-specific patient records belonging to two different hospitals, drawn from the 184 records produced by the locked processing steps. Same-hospital pairs were excluded, because records within one hospital are already grouped by file number and the locked rule links across hospitals only; no blocking and no other exclusion was applied. This gives 15,213 candidate pairs: the 16,836 pairs among 184 records less 1,623 same-hospital pairs. Eleven deterministic linkage specifications were applied to every candidate pair — the locked rule and 10 alternatives that dropped or relaxed the file-number, name, sex and age criteria, including approximate name agreement (edit distance of up to three characters, or agreement on first and last name only) and age differences of up to five years — with links closed transitively. Names were compared after the revision-stage normalisation described in S3 Text , and no probabilistic scoring was used. The analysis used the retained identifier-bearing extract solely for deterministic linkage and verification of data accuracy under the existing institutional approval (section 11), and only aggregate counts are reported ( S3 Text ; S4 Table ).
Encounters belonging to one linked patient and separated by 14 days or fewer were chained into one algorithmically reconstructed care episode . This rule is present in the original submission but is not documented in the pre-Results analysis plan or data lock , and is therefore non-prespecified. It uses patient identity and visit dates only. Counts were recomputed at 7 and 30 days as a sensitivity analysis of the window. The algorithm, its provenance and its behaviour are set out in S4 Text . The retained fields cannot establish whether encounters close in time belong to one bite or to separate bites, so reconstructed episodes are not equated with clinical bite events.
Hospital charts were reviewed at three of the 14 hospitals, selected pragmatically because their clinical records were accessible for detailed review, not by random or representative sampling. Of 37 eligible records, 35 charts were retrieved; all 35 matched their registry record. This confirms only that the intended charts were retrieved; it does not validate the diagnostic coding, the encounter consolidation, the linkage or the episode reconstruction. Abstraction used a predefined structured form and was performed by one author (H.M.T.), who carried out the chart review and the primary abstraction of the predefined clinical variables for the 35 retrieved records. Other authors contributed to subsequent data checking and manuscript preparation, but the original abstraction was not independently duplicated by a second reviewer ; inter-rater reliability is therefore not estimated.
This component is a chart review and clinical characterization, not a diagnostic validation study . No independent, systematically collected reference classification exists across the 35 records, and the abstraction did not capture the information needed to classify each record retrospectively as confirmed snakebite, probable snakebite, dry bite, non-venomous bite or miscoding. We therefore do not impose those categories and estimate no positive predictive value, sensitivity or specificity for the registry case definition. Age in this subset is the age documented at the snakebite presentation and is not interchangeable with the registry age described insection 3. The abstraction framework is given in S5 Text .
Shock, acute kidney injury, coagulopathy, neurotoxicity and compartment syndrome were recorded as documented by the treating team , with no retrospective reclassification from laboratory values. Coagulopathy is a clinician-recorded variable and was not derived from the recorded clotting-time value. The clotting-time field is reported as a recorded clotting time ; the underlying procedure is not documented in the source, so it is not described as a standardised or validated 20-minute whole-blood clotting test, and no analytic definition depends on it.
Throughout the study period a single antivenom product was in use across the participating hospitals. The abstraction recorded whether antivenom was given, the interval to first administration and the number of administrations, but no product name, volume or neutralising potency, so the product was not verified at the level of the individual chart.
No analysis of antivenom indication is reported. The abstraction's documented-indication field was completed for all 17 treated patients and for none of the 14 untreated, so any concordance computed from it is determined by design. In addition, the composition of the indication rule applied in the original analysis is not documented in the retained study materials. Antivenom is therefore reported only as administered or not. Chart documentation cannot establish whether an indication was recorded before administration, so no temporal ordering and no appropriateness, effectiveness or guideline-adherence claim is made. A previously reported composite adverse-outcome variable was near-redundant with its biological components and is not used; components are reported separately. Variable-level completeness is given in S3 Table .
Counts and proportions are reported with exact (Clopper–Pearson) 95% confidence intervals [ 14 ] where a denominator is defined. No prespecified sensitivity analysis existed for the episode rule; the analyses reported as sensitivity are the post-lock linkage alternative, the post-hoc revision-stage linkage specification analysis and the 7- and 30-day episode windows. Which analyses were prespecified, non-prespecified or revision-stage is recorded in S1 Text . Analyses were performed in Python; the computational specification is given in S6 Text , and the code, with a synthetic input fixture, is publicly available at https://doi.org/10.5281/zenodo.23002338 (version 1.0). The code for the revision-stage linkage sensitivity analysis, with a synthetic identifier-format fixture and its aggregate reference output, is deposited as version 1.1 of the same record at https://doi.org/10.5281/zenodo.24015979 ; running it on study data requires the restricted identifier-bearing extract. Reporting follows STROBE [ 15 ] ( S1 Checklist ) and RECORD [ 16 ] ( S2 Checklist ).
The determination states that the Committee “approves temporary use of identifiers solely for deterministic linkage, episode reconstruction, duplicate removal, source-record validation and verification of data accuracy,” that temporary access to identifiable information is permitted only for deterministic record linkage and validation, that direct identifiers shall be removed following completion, and that access remains restricted to authorised investigators. Administrative permissions and data-access authorisations are the responsibility of the participating institutions.
Identifier-bearing source extracts were necessarily used for exact deduplication, encounter construction, patient linkage and temporal episode reconstruction, because those operations depend on the patient file number, the patient name and exact visit dates. The derived dataset used for the reported aggregate results is pseudonymised rather than anonymised : it carries no names, file numbers, dates of birth or exact dates, but retains pseudonymous patient and site identifiers and day offsets. The post-hoc revision-stage linkage sensitivity analysis (section 6) used the retained identifier-bearing extract solely for deterministic linkage and verification of data accuracy under the existing institutional approval and data-custodian arrangements (PMCQu-22-04-2026). This revision-stage use was confirmed as covered by the existing governance framework, and no additional ethical amendment was required. Access was by authorised investigators, the computing environment used for this work was confirmed as compatible with the applicable confidentiality and access requirements, and only aggregate, non-identifiable outputs were retained for reporting and reproducibility. The linkage and verification procedures are now complete; identifier-bearing copies are handled according to the data custodian's instructions and the existing approval. The date on which identifier-bearing copies are finally removed is not documented and is not stated here.
Patient-level data are not publicly deposited. Enquiries or requests concerning potential access may be directed to the Institutional Research Ethics Committee of Palestine Medical Complex, Palestinian Ministry of Health, Ramallah, Palestine, at pmc@moh.ps , quoting reference PMCQu-22-04-2026; each request is considered case by case through the applicable institutional process, and submission of a request does not guarantee access. The dates on which the registry extract and the clinical records were first accessed for research purposes are not recorded in the retained documentation and are not inferred from the 26 April 2026 extraction date. Processing and consolidation are detailed in S2 Text ; the analysis plan and provenance record in S1 Text .
ChatGPT (OpenAI) and Claude (Anthropic) were used during the preparation and revision of this manuscript. Their roles were language and structural editing, methodological and consistency review, code development and checking, reproducibility verification, and assistance with the post-hoc revision-stage linkage sensitivity analysis. That revision-stage work involved processing the restricted identifier-bearing extract within the authorised analysis environment, under the arrangements described insection 11; the outputs reported and deposited are aggregate and non-identifiable. Neither tool performed the original clinical chart abstraction or made clinical decisions, and no other generative-AI tool was used. The authors independently reviewed and verified all AI-assisted outputs against the underlying study materials. The authors retain full responsibility for the analyses, interpretation, accuracy, integrity and final content of the manuscript.
The extract contained 362 rows . 70 rows were identical on every field and were removed, leaving 292 retained rows . Consolidation by hospital, patient file number and visit date produced 265 registry encounters , which resolved into 184 hospital-specific patient records and, after cross-hospital linkage, 172 globally linked patients ( Table 1 ). Applying the 14-day algorithm produced 178 algorithmically reconstructed care episodes ( Fig 1 ; S2 Table ).
https://doi.org/10.1371/journal.pntd.0014897.t001
Left branch, registry reconstruction: 362 extracted rows; 70 rows identical on every field removed, leaving 292 retained rows; consolidation of records sharing hospital, patient file number and visit date giving 265 unique encounters; 184 hospital-specific patient records; 172 linked patients under the locked cross-hospital linkage rule; and 178 algorithmically reconstructed episodes under the non-prespecified 14-day rule. Right branch, nested chart review, drawn from the 172 linked patients: 37 patients were eligible because they had at least one encounter at one of the three chart-review hospitals, 35 charts were retrieved, and 31 were clinically assessable, 4 having died before arrival. The remaining 135 of 172 had no encounter at one of the three chart-review hospitals and were therefore not eligible for the nested chart review; all 172 were documented within the 14-hospital registry.
https://doi.org/10.1371/journal.pntd.0014897.g001
Eight distinct T63 codes appeared across the 265 encounters, and the distribution of subcodes varied across hospitals: the modal code accounted for between 46% and 91% of a hospital's encounters. Two encounters carried species-denominated subcodes; these are treated as coding labels and are not evidence of species identification ( S1 Table ).
Twelve cross-hospital links were accepted under the locked rule, involving 12 of the 14 hospitals across 9 distinct hospital pairs. 12 of 172 patients (7.0%, 95% CI 3.7–11.9) were documented at more than one hospital ( Table 2 ; Fig 2 ; S1 Fig ).
https://doi.org/10.1371/journal.pntd.0014897.t002
Paired horizontal bars show, for each of the 14 hospitals, the number of patients documented and the number of registry encounters. The difference between the bars is the hospital-level repeat-attendance load.
https://doi.org/10.1371/journal.pntd.0014897.g002
Accepting the 13th potential link gives 171 patients, of whom 13 (7.6%, 95% CI 4.1–12.6) were documented at more than one hospital, and episode counts of 183, 178 and 176 at 7, 14 and 30 days. The largest movement in any reported proportion was 0.65 percentage points ( S4 Table ). This bounds the effect of that single potential link and is not evidence that linkage error in general is small.
Across the 11 linkage specifications, the number of linked patients was 172 or 171 ( S4 Table ). The four specifications that required file-number agreement, including file-number agreement alone, each reproduced the 12 primary links exactly and gave 172 patients. The seven specifications that did not require file-number agreement each gave 171 patients; in every case the single additional link was the post-lock pair described above. No specification removed a primary link, and no specification added any other pair. Among the 15,213 candidate pairs, the 12 cross-hospital pairs sharing a file number all had concordant name and sex. This is internal-consistency evidence only and does not establish the absence of false-positive links. These results support robustness to the observed linkage-rule choices within the available variables; they do not establish absolute linkage accuracy and are not external validation.
The 265 registry encounters were grouped into 178 reconstructed episodes : 87 encounters were additional to the first encounter of their episode. Expressed as a ratio, encounters were 48.9% more numerous than reconstructed episodes. Of 93 inter-encounter intervals, 87 were 14 days or fewer; the six intervals that created episode boundaries were 15, 34, 112, 125, 364 and 503 days. Because the rule chains transitively, three episodes span more than 14 days (15, 18 and 21 days), while 113 episodes span a single day. Eleven of 178 episodes include encounters at two hospitals, and 6 of 172 patients contributed more than one episode. Across alternative windows the counts were 183 (7 days), 178 (14 days) and 177 (30 days) : episode counts varied modestly across alternative algorithmic windows ( S4 Text ).
Records were unevenly distributed: 18 of 70 hospital-years contained no record , and the number of hospitals contributing any record ranged from 8 to 12 across the five years ( Fig 3 ; S5 Table ). Temporal patterns therefore reflect both event occurrence and which hospitals recorded anything in a given year, and cannot be read as incidence.
(A) Encounters by calendar year, with the number of hospitals recording at least one encounter annotated above each bar. (B) Encounters by calendar month pooled across 2021–2025, with May to October shaded. Because 18 of 70 hospital-years contain no record, these distributions reflect both event occurrence and hospital recording, and are not incidence.
https://doi.org/10.1371/journal.pntd.0014897.g003
Of the 172 linked patients, 137 (79.7%, 95% CI 72.9–85.4) were male . Age, as calculated by the hospital information system from date of birth at the extraction date, had a median of 30 years (IQR 18.8–48.0; range 1–82); 36 (20.9%, 95% CI 15.1–27.8) were under 18 years and 14 (8.1%) were 65 or over at that reference date . Because the reference date follows every encounter, these figures describe age at extraction and not at presentation: the proportion under 18 may underestimate, and the median and the proportion aged 65 or over may overestimate, the corresponding values at presentation. Age and sex were complete.
Of 37 eligible records, 35 charts (94.6%, 95% CI 81.8–99.3) were retrieved ; all 35 matched their registry record. Four patients were documented as dead before arrival and are not assessable for in-hospital findings, leaving 31 clinically assessable patients . In this subset the age documented at presentation had a median of 29 years (IQR 21–45; range 7–81), and 2 of 35 were under 18; because registry age refers to a different point in time, the two age distributions are reported separately and not compared ( Table 3 ).
https://doi.org/10.1371/journal.pntd.0014897.t003
Among the 31: a documented systemic manifestation appeared in 12; clinician-recorded coagulopathy in 10; shock in 3; acute kidney injury in 2; neurotoxicity in 2; and compartment syndrome in 2. Antivenom was administered to 17, intensive care used for 11, surgery performed in 3, and finger amputation in 1. There was no in-hospital death (0 of 31; upper 95% bound 11.2%). Documented manifestations are reported as strata rather than diagnostic categories: 12, 2, 2 and 15 of the 31 by stratum, with the 4 patients dead on arrival not assessable ( Table 4 ; Fig 4 ).
https://doi.org/10.1371/journal.pntd.0014897.t004
Forest plot of proportions with exact (Clopper–Pearson) 95% confidence intervals; the denominator is the 31 clinically assessable patients. Individual manifestations, treatment and in-hospital death are shown.
https://doi.org/10.1371/journal.pntd.0014897.g004
No indication analysis is reported: the abstraction's documented-indication field was completed for all 17 treated patients and for none of the 14 untreated, so any concordance computed from it is determined by design. Three of the 17 treated charts recorded structured swelling as absent while free text described progressive swelling — a discrepancy between two fields of the same chart. No inference about clinical appropriateness, guideline adherence, product performance or the order of events is drawn.
Median recorded clotting time was 14.5 minutes where coagulopathy was recorded and 7.0 minutes where it was not. Two values of 20 minutes or more (23 and 28) occurred without recorded coagulopathy, and 5 of 10 patients with recorded coagulopathy had values below 15 minutes. Coagulopathy was therefore not a function of the recorded clotting time ( S3 Table ).
Four operations applied to one hospital extract produced four different quantities: 362 rows, 265 registry encounters, 172 linked patients and 178 reconstructed episodes. The same source extract yielded substantially different counts after exact deduplication, encounter consolidation, patient linkage and algorithmic episode reconstruction. The contribution of this study is to make each step explicit, to report what each produces, and to state what none of them establishes.
The counts reported here describe hospital-recorded care. They are not estimates of snakebite incidence or burden. Every unit in the reconstruction — row, encounter, linked patient or episode — exists only because a person reached a public hospital, was documented and was assigned a retrievable T63 code. Where hospital and population sources have been compared directly, the difference has been large. In one Sri Lankan district, the population incidence of hospitalised snakebite was 158 per 100,000, against community survey estimates of 499 per 100,000, although the incidence of envenoming did not differ significantly between the two sources [ 7 ]. A national community survey in Sri Lanka attributed part of the shortfall in hospital statistics to victims seeking traditional treatment [ 8 ], and in India routine reporting captured only about 10% of the expected hospital-based snakebite deaths [ 9 ]. These comparisons cannot be transferred to the West Bank, where no community survey exists, and we do not use them to scale our counts. They show why a hospital-derived count, however carefully reconstructed, is an incomplete and selected measure of hospital-recorded activity, conditioned on care-seeking, documentation and coding, rather than an estimate of population occurrence. Because the diagnostic coding was not validated, such a count may also include records that were not snakebites, so it cannot be treated as a lower bound on occurrence either. The observation that all four documented deaths in the chart-reviewed subset preceded arrival points the same way: deaths that do not reach hospital lie outside a hospital-based count by construction, and the registry contains no outcome field, so network mortality is unmeasured rather than low.
Linkage converts an administrative count into a person count, and some level of linkage error or uncertainty is almost always inevitable when administrative datasets are linked [ 17 ]. Deterministic techniques are often designed to minimise false links [ 17 ], which places the residual risk on missed links, and errors in either direction can bias results derived from linked data [ 17 ]. No external authoritative identifier was available; absolute linkage error remains unquantifiable. A national identification number exists in the source system but was not included in the analytic extracts, so the linkage could not be compared with a gold standard. Of the three approaches to evaluating linkage quality described by Harron and colleagues [ 13 ], the one feasible here was sensitivity analysis of how strongly the linked-patient count depended on plausible changes in the linkage procedure; without a reference standard, this analysis cannot quantify false or missed links. Across 11 deterministic specifications, including approximate name matching and relaxed age criteria, the linked-patient count was 172 or 171, and the only additional link under any specification was the single post-lock pair already reported. These results support robustness to the observed linkage-rule choices within the available variables; they do not establish absolute linkage accuracy and are not external validation. All alternative specifications could fail to detect the same person if both the identifier and the name differed substantially between hospitals, so the sensitivity analysis does not exclude missed links. The absence of any file number shared by discordant names or sexes is likewise internal-consistency evidence and does not demonstrate the absence of false-positive links. A count of candidate pairs considered during the original linkage was not retained; the candidate-pair space analysed during revision was reconstructed de novo and does not recover the original adjudication.
Administrative systems record contacts with services, and the unit they record is rarely the unit of clinical interest. Large hospital databases address this explicitly. In English Hospital Episode Statistics, a spell is an uninterrupted inpatient stay at one hospital that does not include transfers between hospitals, and continuous inpatient spells must be derived to follow a patient across hospitals [ 18 ]. In Manitoba, failing to account for inter-hospital transfers when constructing episodes of inpatient care overstated the number of episodes [ 19 ]. Those systems differ from ours in data structure, time resolution and clinical setting, and their figures are not comparable with ours; the parallel is methodological. In each case the count depends on an aggregation rule that the records do not themselves supply. Here, consolidating departmental rows into encounters and grouping encounters into reconstructed episodes changed the count at each step. The 14-day rule was non-prespecified and has no external referent, and it determines where one reconstructed episode ends and the next begins; reconstructed episodes cannot be equated with clinically confirmed bites . The 7/14/30-day comparison evaluates numerical dependence on the window, not clinical validity: every count comes from the same rule applied to the same dates. The retained registry data cannot distinguish follow-up, referral or readmission encounters from new bite events , which is precisely the ambiguity the episode algorithm resolves by assumption rather than by evidence.
Three features of these records limit their use for surveillance beyond the reconstruction itself. T63 coding does not itself confirm snakebite or envenoming ; it is the ascertainment criterion, and using coded records to confirm the same diagnosis would be circular. Coding practice varied between hospitals, and 18 of 70 hospital-years recorded no encounter, so year-to-year variation mixes occurrence with recording. The clinical component is narrow: chart review was restricted to three hospitals , 137 of the 172 linked patients had no chart reviewed, no independent reference classification exists across the 35 records, and independent duplicate abstraction was not completed, so no positive predictive value or inter-rater reliability is estimated and the clinical findings cannot be generalised to the 14-hospital extract , still less to the population. Each of these is a property of how the records are made rather than of this analysis, and each points to a specific change in recording practice — a documented case definition, an identifier for the period of care, documented identity resolution across facilities and a minimum structured clinical dataset — set out as a proposal in S7 Text .
What the study offers is a method and a set of bounded observations: the same routine extract gives substantially different counts depending on whether rows, encounters, linked patients or reconstructed episodes are counted; the rule at each step can be written down and tested for sensitivity; and the linked-patient count did not change materially across the linkage specifications that the available variables allow. What it does not establish is equally specific. It does not estimate snakebite incidence or burden, does not confirm that any coded record was a snakebite or an envenoming, does not validate the linkage or the episode rule against an external standard, and does not measure mortality across the network. Reported this way, the counts are useful for surveillance planning and for anyone reading figures derived from similar systems; reported as a burden estimate, they would overstate what the records can establish.
Case ascertainment rests on diagnostic coding alone; no registry record can establish that a snakebite occurred, that venom was injected, or that a species was identified. Linkage relied on fields that cannot validate themselves, and no reference standard was available in the analytic extracts. No external authoritative identifier was available; absolute linkage error remains unquantifiable. The post-hoc revision-stage sensitivity analysis showed that the linked-patient count was 171 or 172 across 11 linkage specifications, but all of those specifications could fail to detect the same person if both the identifier and the name differed substantially between hospitals; the sensitivity analysis therefore does not exclude missed links. Candidate-pair information from the original process was not retained. The episode rule was not prespecified and cannot be checked against clinical events. Registry age refers to the extraction date rather than to each encounter, so age-based descriptions of the linked cohort describe that reference point. Coverage across hospital-years was uneven, which precludes temporal or incidence inference. For the complete calendar years 2022–2024, Palestinian Ministry of Health annual reports recorded 193, 183 and 214 reported snake bites at MoH hospitals in the West Bank [ 20 – 22 ], whereas the T63-coded hospital information system extract analysed in this study yielded 58, 37 and 59 consolidated encounters. These figures arise from differently constructed routine-data products, and the available documentation does not establish that their case definitions, counting units or ascertainment pathways are equivalent. We therefore treat the two sources as non-interchangeable and do not use the annual-report figures to validate, recalibrate or scale the study counts. The nested chart review covered three hospitals, was performed by a single abstractor without duplicate review, and characterises clinical documentation rather than validating any registry procedure. Chart documentation cannot establish whether antivenom indications were recorded before administration. The recorded clotting time is of unverified procedure. Two provenance details are not documented in the retained study materials: the normalisation procedure applied to names in the original linkage, and the date on which identifier-bearing copies are finally removed. Neither affects a reported result, and neither is assumed in the text.
Routine hospital records can be reorganised into a transparent sequence — rows, encounters, hospital-specific records, linked patients, reconstructed episodes — and each step changes the count materially. The resulting figures are hospital-recorded, registry-derived reconstructions, dependent on a deterministic linkage rule and a non-prespecified episode algorithm; they are not estimates of snakebite incidence or burden, and they are not clinically validated counts of snakebites or envenomings. Reported this way, routine records can support surveillance planning and can be audited by others; reported as a burden estimate, they would overstate what the underlying records can establish.
https://doi.org/10.1371/journal.pntd.0014897.s001
https://doi.org/10.1371/journal.pntd.0014897.s002
https://doi.org/10.1371/journal.pntd.0014897.s003
https://doi.org/10.1371/journal.pntd.0014897.s004
https://doi.org/10.1371/journal.pntd.0014897.s005
https://doi.org/10.1371/journal.pntd.0014897.s006
https://doi.org/10.1371/journal.pntd.0014897.s007
https://doi.org/10.1371/journal.pntd.0014897.s008
https://doi.org/10.1371/journal.pntd.0014897.s009
https://doi.org/10.1371/journal.pntd.0014897.s010
https://doi.org/10.1371/journal.pntd.0014897.s011
https://doi.org/10.1371/journal.pntd.0014897.s012
https://doi.org/10.1371/journal.pntd.0014897.s013
https://doi.org/10.1371/journal.pntd.0014897.s014
https://doi.org/10.1371/journal.pntd.0014897.s015
The authors thank the medical records and quality departments of the participating hospitals.
For more information about PLOS Subject Areas, click here .
Is the Subject Area "Snakebite" applicable to this article? Yes No
Is the Subject Area "Charts" applicable to this article? Yes No
Is the Subject Area "Coagulopathy" applicable to this article? Yes No
Is the Subject Area "Epidemiology" applicable to this article? Yes No
Is the Subject Area "Palestinian territories" applicable to this article? Yes No
Is the Subject Area "Algorithms" applicable to this article? Yes No
Is the Subject Area "Hospitals" applicable to this article? Yes No
Is the Subject Area "Surgical and invasive medical procedures" applicable to this article? Yes No
PLOS is a nonprofit 501(c)(3) corporation, #C2354500, based in California, US
Loading comments…