Bracketing the Ingestion Clock: Measuring and Reducing Caregiver Recall Error in Pediatric Magnet Ingestion

When a toddler swallows several small magnets, the magnets can attract each other through two loops of bowel or two folds of stomach. They pinch the tissue between them, cut off its blood supply, and over hours to days erode through it. The emergency team usually decides what to do from two inputs: an abdominal X-ray, and a parent's answer to "when did this happen, and how many did they swallow?" Both inputs are unreliable. Most of these children look well on arrival: Sarmini et al. found 66.3% of pediatric foreign-body ingestions asymptomatic at presentation, while magnets needed active removal in 35% of cases. A flat X-ray can also make beads clamped across separate loops look like one harmless chain, as in the infant case of Nakib et al.. The parent's account is a memory reconstructed under fear and guilt, usually about an event nobody watched. No study has measured how wrong that account is.

This note proposes the first ground-truthed measurement of caregiver recall error in pediatric ingestions, with magnets as the primary subgroup. The study then tests one concrete fix: a structured "bracketing" interview that asks for the last time the magnet set was seen complete and the first time pieces were noticed missing, each tied to a daily-routine landmark. It returns an interval rather than a single guessed time. The study also asks a question that could make all of this unnecessary: does the history change management at all, compared with deciding from imaging alone?

Core Claim

The claim has three parts. First, caregiver point estimates of ingestion time and magnet count contain systematic error, not just random error. The predicted direction is that reported times are too recent and reported counts too low. The error should be largest when ingestion was unwitnessed and the caregiver was present but not watching, the "supervision paradox" that Wang et al. documented in 67.71% of caregiver-present childhood injuries. Second, asking for an interval anchored to routine landmarks, framed without blame, will cover the true ingestion time more often than a single point estimate does. Third, and this is the claim that must survive, the improvement matters only if it changes triage decisions. The study is built so that a negative answer on the third point is just as publishable as a positive one, and it would redirect the work.

Prior Work and Its Role

Two kinds of prior work motivate the expectation of directional bias. One shows that current emotional state rewrites the remembered past. Wu et al. found that people in depressive episodes over-report past adversity. Pfeiffer et al. and Cataldo et al. found that current distress and current attachment reshape memories of family climate and parental bonding. Lyons showed that people retrospectively inflate their health before an injury. The other kind comes from injury epidemiology: Wang et al. note that guardians under-report and distort injury details out of reluctance and emotional trauma. None of these studies concerns time-critical ingestion histories, and none had ground truth for the event itself.

Two papers supply the method. Guo et al. compared mothers' self-reported acetaminophen use with maternal and newborn biomarkers, and that comparison against an objective anchor is the design we reuse. Liu et al. and Zhang et al. show that error grows with the length of the recall interval. That motivates recording elapsed time from ingestion to interview as a covariate.

Two papers constrain the design. Nissen et al. show that even prospective app logging badly undercounts unless users are highly engaged: back-pain prevalence was 44% among all users versus 93% among those logging at least five months. Prospective household logging is therefore not a realistic route to ground truth for rare accidents. CT-based data generation for foreign object detection on a single X-ray projection formalizes how a single projection flattens 3D arrangement into overlapping shadows. This is why X-ray position can serve only as a coarse clock and a consistency check, never as a precise ingestion time.

Finally, Sarmini et al. and Nakib et al. define the clinical regime. Both take the caregiver's history as given and never measure its accuracy. That gap is what this study fills.

Research Program

The study is a prospective, multicenter observational study of children presenting to pediatric emergency departments with suspected radio-opaque ingestions. It is nested with a cluster-randomized comparison of two interview protocols. Magnets are the primary analysis subgroup. All radio-opaque ingestions are enrolled for estimating recall error, because the memory mechanism is shared across object types and timing matters acutely for button batteries too. Pooling is also necessary on sample-size grounds. At the Sarmini center, 11.8% of 338 ingestions over roughly fifteen months were magnets, about 40 cases. A single site cannot power a magnet-only analysis, so the study needs several centers and a recruitment window of at least two years.

Ground truth for ingestion time has three tiers, and every analysis is reported separately by tier. Tier one is a time-stamped trace: home-camera footage, a timestamped photo or message referring to the event, or a timestamped call to a poison line. Tier two is an independent account from a second adult who witnessed the event, interviewed separately and before the first caregiver's account is shared. Tier three is a radiographic transit stage (stomach, small bowel, colon), compared against the reported interval; it serves only as a plausibility check. Ground truth for magnet count is the number retrieved at endoscopy or surgery, or the number on serial imaging for conservatively managed cases. Each child also gets a structured record of intra-procedural mucosal injury grade. This provides a tissue-level signal of how long compression lasted, which can be cross-checked against the reported times.

The two interview arms are the current free-form triage question ("when did they swallow it, how many?") and the bracketing interview. The bracketing interview elicits, in a fixed order, the last time the set was seen complete with its count, the first time anything was noticed missing, and the routine events around each (meals, naps, daycare pickup). It uses explicitly non-blaming wording. Arms alternate by calendar block within site, so staff do not mix scripts. The interview is administered in addition to normal clinical care and never replaces it. Treating clinicians see the clinical history they would normally take.

The primary measures for time are signed error of the point estimate in hours against tier-one or tier-two truth, and coverage of the bracketed interval. The primary measure for count is signed count error. Pre-specified covariates are witnessed status, caregiver identity (parent versus grandparent, given the higher injury risk under grandparental care in Wang et al.), a brief distress scale, disclosure of a supervision lapse, and the delay from ingestion to interview.

Baselines

The naive reference point is the caregiver's unprompted point estimate taken at face value. This is what the Sarmini and Nakib case series implicitly do. It is run on the same children under the standard-interview arm. Any intervention must reduce absolute timing error and raise coverage relative to it. No reported result was found on Lacuna for the accuracy of caregiver-reported ingestion time, so the size of this error is itself the first finding.

The strongest existing statistical competitor is regression calibration, the classical measurement-error correction. It fits a model mapping reported time and covariates to true time on the tier-one/tier-two validation subset, then applies that model to everyone. It is the strongest competitor because it needs no change to clinical practice, only data, and it directly targets systematic bias. It gets the same covariates and the same validation subset, and is evaluated with cross-validation across sites, so that it is never fit and scored on the same center. The bracketing interview beats it only if bracketing's coverage and error on held-out sites are better than the corrected point estimates.

The imaging-only reference decides triage from X-ray count and location alone, ignoring history. This baseline answers whether history matters at all, and it is evaluated on decisions rather than on timing.

Two ablations isolate the key ingredient. One uses landmark bracketing with neutral rather than non-blaming wording. The other uses non-blaming wording with a single point question. Both are administered in later calendar blocks at a subset of sites. If non-blaming wording alone captures most of the gain, the mechanism is guilt-driven distortion, not memory decay, and the cheaper intervention wins.

For context only, the one recall-correction system in the file with a reported number is FLARE in Chen et al.. It reports 89.5% accuracy and an F1 of 0.893 on predicting wildfire evacuation decisions, against 69.7% for logistic regression. That is a different task, decision prediction rather than fact recovery, and is not a target here.

Why not just…?

Why not just ignore the history and treat every multi-magnet X-ray as urgent? This is the most serious threat. If management of multiple magnets is effectively history-independent, recall correction cannot change magnet outcomes. The decision-impact experiment below tests this directly. If imaging-only triage agrees with ground-truth-informed triage in essentially all magnet cases, we say so. The magnet-specific claim then shrinks to two points where imaging cannot help: whether a caregiver brings an asymptomatic child in at all, and cases where superposition makes count or configuration ambiguous, as in Nakib et al.. Button batteries, where elapsed time drives urgency, would become the main beneficiary.

Why not just apply standard measurement-error correction to the reports you already get? This is a real competitor, and it is run as the main baseline. It has a structural limit. It corrects average bias given covariates but cannot recover information the caregiver never retrieved. Bracketing tries to elicit additional anchored memories. If cross-site regression calibration matches bracketing, the field should adopt the correction model, because it costs nothing at the bedside.

Why not just use a language model to reinterpret the caregiver's narrative, as FLARE did for evacuation surveys? FLARE in Chen et al. predicts what people did from biased recollections. It does not recover when an unobserved physical event occurred. Without ground-truth timestamps, a model would only learn to reproduce plausible-sounding narratives. A text model could be layered on later, once the tier-one data from this study exist to train and test it.

Why not just log exposures prospectively at home? Nissen et al. show heavy undercounting outside a highly engaged minority. Unwatched accidents during leisure at home, the dominant pattern in Wang et al., are by definition not logged.

Proposed Experiments and Decisive Checks

The first experiment estimates the error distribution. On the standard-interview arm, we estimate the signed timing and count error against tier-one and tier-two truth, stratified by witnessed status, caregiver identity, and supervision-lapse disclosure. The directional hypothesis is falsified if unwitnessed, lapse-disclosing cases show no shift toward too-recent times and too-low counts relative to witnessed cases. A pre-check compares tier-one cases with tier-two-only cases on demographics and object type, because households with cameras may differ systematically.

The second experiment compares interview arms. Bracketing is compared against free-form reporting and against regression-calibrated point estimates on interval coverage and absolute error, using held-out sites. The two ablations are analyzed the same way.

The third experiment is decisive: decision impact. A blinded panel of pediatric gastroenterologists reviews each magnet and battery case under four information sets: X-ray only; X-ray plus history as reported; X-ray plus bracketed history; and X-ray plus ground truth. The outcome is the rate of under-triage relative to the ground-truth decision. Under-triage means choosing observation where the ground-truth panel chose retrieval. Margins are pre-registered before unblinding. The proposal succeeds only if bracketed history reduces under-triage relative to reported history and is not dominated by X-ray-only triage.

A secondary check correlates ground-truth elapsed compression time with endoscopic mucosal injury grade in magnet cases. This gives the first in-patient data on how fast pinched tissue progresses, which Sarmini et al. and Nakib et al. could not provide because their timelines rested on uncorrected recall.