Child abuse medical records: sentinel injuries, abusive head trauma and the workup

There is no single chart. A physical abuse evaluation pulls records from every place the child has been seen, to ask whether the injuries and the explanation fit an accident, a medical condition or an inflicted injury.

The literature is younger than many assume. C. Henry Kempe and colleagues published "The battered-child syndrome" in JAMA in 1962. Most of the screening rules and imaging protocols here date from the last 20 years, so older records may predate them.

The AAP's 2015 clinical report on the evaluation of suspected child physical abuse lists the physician's roles: identifying abused children, reporting to the child protection agency, coordinating with other professionals and "providing court testimony when necessary." Each role leaves a document.

A prosecutor reads the file for injury, timing and access. Defense counsel reads it for the alternative the workup missed and for anchoring, an early label that shaped every later note. Family court counsel reads it for safety, on a faster clock. Civil counsel reads it for failure to diagnose and failure to report (chapter 7).

Sentinel injury

A minor injury, such as a bruise or mouth injury, in an infant, suspicious because the baby could not cruise or the explanation was implausible, found before a more serious abusive injury.

Abusive head trauma (AHT)

The AAP's term since 2009 for inflicted injury to the head and its contents, from shaking, impact or both.

Skeletal survey

A standardized set of separate x-rays of the whole skeleton, to find fractures that cannot be seen or felt.

Precruiser

A baby not yet pulling up and walking along furniture. Bruising is rare at this stage.

Child protection team

A hospital group, usually led by a child abuse pediatrician, that consults on suspected abuse. Its notes are often filed apart from the treating team's.

In many serious infant abuse cases, the most important record is from a visit weeks before the admission, where a small injury was seen, charted and explained. Sheets and colleagues defined it in a 2013 Pediatrics study of 401 infants under 12 months evaluated by a hospital child protection team, "Sentinel injuries in infants evaluated for child physical abuse." They defined it as "a previous injury reported in the medical history that was suspicious for abuse because the infant could not cruise, or the explanation was implausible."

In the abused group, 80% of sentinel injuries were bruises and 11% were intraoral injuries. 66% happened before 3 months of age and 95% at or before 7 months. The abstract adds that medical providers were reportedly aware of the sentinel injury in 41.9% of cases, without stating the denominator; check the full text before quoting it.

Mobility is the first fact to find

The definition turns on whether the baby could cruise. That comes from a 1999 community study by Sugar and colleagues, "Bruises in infants and toddlers: those who don't cruise rarely bruise," which examined 973 children under 36 months at well-child visits.

The authors concluded that bruises in infants under 9 months not yet beginning to ambulate "should lead to consideration of abuse or illness as causative." The word "illness" is theirs.

How to hunt for a sentinel injury

1. List every encounter from birth. Well-child, urgent care, emergency, telehealth, advice lines, pharmacy.
2. Read the skin and mouth exam in each. Look for bruise, mark, discoloration, petechiae, a torn frenulum or blood in the mouth, and any body diagram.
3. Record age and mobility on that date, from the chart, not from later testimony.
4. Record the explanation, who gave it and the plan: survey, consult, report or nothing.
5. Check portal messages and photos. They are part of the record and often missing from a standard production.

Defense teams run the same hunt: a bruise that was examined, explained and followed up may undercut a theory of repeated abuse.

TEN-4-FACESp is a bruising clinical decision rule. It tells a clinician when bruising in a young child warrants further evaluation for abuse. It does not diagnose abuse, and its authors say so.

Pierce and colleagues derived the original TEN-4 rule in a 2010 pilot study of 95 children admitted to intensive care for trauma: bruising on the torso, ear or neck up to age 4, or anywhere on an infant under 4 months, with 97% sensitivity and 84% specificity. The team then refined the rule and validated it internally (bootstrap resampling) at 5 urban children's hospital emergency departments, published in JAMA Network Open in 2021. They screened 21,123 children, enrolled 2,161 under 4 with bruising, recorded bruising in 34 body regions, and had an expert panel classify 410 as abuse and 1,713 as nonabuse. The refined rule was 95.6% sensitive and 87.1% specific. Its positive predictive value was 63.9%: about 1 in 3 children who screened positive had been classified as nonabuse.

The rule is only as good as the exam behind it.

Fractures in infants are often invisible, and a healing rib fracture may produce no swelling at all. The AAP's 2009 statement on diagnostic imaging of child abuse gives imaging 2 jobs in 1 sentence: "to identify the extent of physical injury when abuse is present and to elucidate all imaging findings that point to alternative diagnoses." A survey is as much a search for a medical explanation as a search for injury.

The technical standard is the ACR and SPR Practice Parameter for the Performance and Interpretation of Skeletal Surveys in Children, revised 2021 and amended 2023. For suspected abuse, each region "should be imaged with a separate radiographic exposure," and:

"A single radiograph (babygram) of the entire infant should not be performed."

The parameter says "a second interpretation by a pediatric radiologist may add value." It lists findings that, "particularly in children younger than 2 years old, suggest a need to consider child abuse and to perform a skeletal series," including bruising, burns, single unexplained fractures, and retinal and intracranial hemorrhages. It calls classic metaphyseal lesions and posterior rib, scapular, spinous process and sternal fractures "highly specific," and warns that "knowledge of variants and simulators of traumatic findings" is necessary.

The follow-up survey

The ACR says a follow-up survey "may be indicated in the setting of nonaccidental injury" but sets no interval. A 2017 review in the Journal of the Belgian Society of Radiology puts it at 10 to 14 days after the first survey, when a fracture invisible on day 1 may show healing. Harper and colleagues measured what the repeat finds, using data from 20 US child abuse teams, published in Pediatrics in 2013.

The last number matters to the defense: a repeat survey can take a suspected fracture off the table. 61% of children with a first survey in that network had no follow-up survey.

Siblings and household contacts

In a 2012 study by Lindberg and colleagues, contacts of abused children were screened by age: exam under 5, exam and skeletal survey under 24 months, and neuroimaging added under 6 months. Surveys found an abusive fracture in 16 of 134 contacts under 24 months (11.9%), none with a finding on exam. Twins had an odds ratio of 20.1 compared with nontwin contacts, so a case may involve several children's charts.

No part of child abuse litigation is more contested than infant head injury. Both sides read the same CT and MRI, eye exam, histories, labs and resuscitation record. What differs is how the literature is read.

From "shaken baby syndrome" to abusive head trauma

In 2009 the AAP Committee on Child Abuse and Neglect issued "Abusive head trauma in infants and children." It said that "although shaking an infant has the potential to cause neurologic injury, blunt impact or a combination of shaking and blunt impact cause injury as well," and recommended that pediatricians:

"embrace a less mechanistic term, abusive head trauma, when describing an inflicted injury to the head and its contents."

The 2020 revision warns that failure to recognize AHT "at any step in the process, from medical diagnosis to child protection and legal decision-making, can place children at risk," and calls itself "not a comprehensive exposition of the science." In the chart, word choice is evidence: "shaken baby" commits to a mechanism, "concern for nonaccidental trauma" does not.

The 2 positions, in their own words

The debate often centers on the "triad" of subdural hemorrhage, retinal hemorrhages and encephalopathy, and how much weight those findings carry in concluding that an infant was injured by another person.

Both documents are narrower than they are often quoted. The review asked about the triad's accuracy for shaking; the consensus addresses AHT as a whole, including impact. Many disputes are about 1 finding, not all 3.

What the courts have said

In Cavazos v. Smith, 565 U.S. 1 (2011), a per curiam opinion reversed a federal appeals court that had overturned a state conviction in an infant shaking case, holding that the jury's verdict was not irrational and that federal courts owed deference to the state court. It added: "Doubts about whether Smith is in fact guilty are understandable." Justice Ginsburg, joined by Justices Breyer and Sotomayor, dissented, writing that doubt has increased in the medical community "over whether infants can be fatally injured through shaking alone," quoting State v. Edmunds, 2008 WI App 33. The case turned on the standard of review, and both sides cite it.

In federal court, admissibility runs through Federal Rule of Evidence 702, amended December 1, 2023 to require the proponent to show it is "more likely than not" that the testimony meets the rule. State rules vary.

Who the experts are

Child abuse pediatrics is a board-certified subspecialty, announced by Block and Palusci in 2006 as "a new pediatric subspecialty." The first certifying exam was given in November 2009, and 184 of 216 candidates passed. Training is a 3-year fellowship after pediatric residency. Check who evaluated the child and which consults are in the file, with dates and reasoning.

The AAP's 2014 report on evaluating children with fractures states the stakes for both sides: "Although the consequences of failing to diagnose an abusive injury in a child can be grave, incorrectly diagnosing child abuse in a child whose fractures have another etiology can be distressing for a family." Its 2022 report on bleeding disorders in suspected child abuse makes 2 points a reviewer should quote together: "not every child with bruising/bleeding that may raise a concern for abuse requires an evaluation for bleeding disorders," and "bleeding disorders cannot be ruled out solely on the basis of patient and family history, no matter how extensive."

The table is not an exhaustive differential. Which tests a child needed is for the experts, and the consensus position on alternative mechanisms is itself contested (chapter 5).

Anchoring, seen from both sides

For the prosecution, anchoring can look like an accidental history accepted without testing. For the defense, it can look like "NAT" written at triage, before imaging or labs, and copied forward while the differential went untested. Either way, date the first appearance of every label, note who wrote it and what evidence existed then.

The Child Abuse Prevention and Treatment Act (CAPTA) is a federal grant statute. It does not itself order a doctor to report; it conditions funding on each state having a reporting system. Under 42 USC 5106a(b)(2)(B), a state must have a law or program that includes:

"provisions or procedures for an individual to report known and suspected instances of child abuse and neglect, including a State law for mandatory reporting by individuals required to report such instances"

The same list requires immunity "for individuals making good faith reports," reaching those who assist "including medical evaluations or consultations" (clause (vii)). CAPTA's definitions, in the notes to 42 USC 5101, set a floor: "at a minimum, any recent act or failure to act on the part of a parent or caretaker, which results in death, serious physical or emotional harm, sexual abuse or exploitation," or which "presents an imminent risk of serious harm."

Everything else is state law: who must report, the level of suspicion, the deadline and the penalty. Read the statute where the care happened, as it stood on the date of care.

What HIPAA allows

45 CFR 164.512(b)(1)(ii) permits disclosure to "a public health authority or other appropriate government authority authorized by law to receive reports of child abuse or neglect." And under 45 CFR 164.502(g)(5), a covered entity may decline to treat a person as the child's personal representative if it reasonably believes the child "has been or may be subjected to domestic violence, abuse, or neglect by such person," and decides that doing so is not in the child's best interest. When both conditions are met, a parent's request for the child's records can be refused.

Failure to diagnose and failure to report

The leading early case is Landeros v. Flood, 17 Cal. 3d 399 (1976). The complaint alleged that a physician and hospital failed to diagnose battered child syndrome in an infant, who was returned to her mother's custody, and that the abuse resumed. The trial court dismissed; the California Supreme Court reversed, letting the claim go forward.

In a 1999 JAMA study of 173 children under 3 with abusive head injuries, Jenny and colleagues found that 54 (31.2%) had been seen by physicians after the injury without the diagnosis being made. Mean time to correct diagnosis was 7 days, and 4 of 5 deaths in that group might have been prevented by earlier recognition. The authors also noted that diagnosis "can be difficult in the absence of a history." Each side cites the half that helps it.

Whether a child can sue a mandated reporter who failed to report is a state law question with no national answer. The good-faith immunity CAPTA requires shields reporters from suits by families later cleared, within limits each state sets.

Before anyone reads for meaning, the file needs 1 source list, 1 page numbering system and 1 timeline.

The route depends on who asks: statutory authority, subpoena or court order, often under a protective order, and state law may add protections for abuse records. The basics are in how to get medical records for a lawsuit. If a late entry matters, the hospital's own EHR audit trail is a separate request (EHR audit trails in medical malpractice).

1. Build the source list: facility, dates requested and received, format, page count.
2. Separate the children before numbering. Sibling pages in the index child's file are common after contact screening.
3. Number every page once, before sorting, so every cite survives reordering.
4. Remove exact duplicates, keep near-duplicates. A copy with 1 added line may be an addendum; it belongs next to the original.
5. Compute age at every encounter from the date of birth. The sentinel and TEN-4-FACESp analyses depend on it.
6. Run 2 tracks: what happened to the child, and who knew what when.

• Every prior encounter since birthIncluding urgent care, telehealth and advice lines
• Birth hospital recordsDelivery, newborn exam and imaging
• Growth charts and milestonesMobility on each date
• Portal messages and photosOften outside a standard production
• EMS run sheetFirst history and scene findings
• All imaging in DICOM with reportsIncluding the follow-up survey
• Retinal photographsFrom the device system if needed
• Consults and send-out resultsHematology, genetics, endocrinology
• Child protection note and photographsAsk for the note type by name
• Social work notesReport details and safety plans
• Agency report recordTime, report number, reporter
• Siblings' screening recordsSeparate authority, separate sources

0 of 12 checked

The file arrives as 1,900 pages from 7 sources. The reviewer builds the source list, numbers pages, computes age at each encounter and lays every visit on 1 timeline.

A prosecution reviewer sees a bruise at 6 weeks in a baby who could not roll, not followed up, and a history that changed between triage and the physician note. A defense reviewer sees rib findings called "possible" with no repeat survey, a missing hematology result and missing retinal photographs. Both need the 12-week portal photo. The timeline makes each reading checkable.

The work in chapters 8 and 9 is sorting thousands of pages, which suits AI medical record review. A large language model (LLM) using clinical natural language processing can find every skin exam, stated history and pending test faster than a person paging through PDFs. The deciding evidence, though, is often not text: photographs, x-rays, CT, MRI and body diagrams.

Where it fails

• Images. A model reading a report sees words about an image. It cannot read a survey, judge a bruise photo or date a fracture.
• Hedges. AI summaries that drop "possible" turn "possible healing rib fracture" into "rib fracture." Check every summarized finding against the source words.
• Labels that spread. If triage wrote "NAT" and 40 notes copied it forward, a model counting mentions treats it as well supported. The first mention matters more than the count.
• OCR and handwriting. Optical character recognition (OCR) and handwriting recognition misread scanned dates, and a 1 cm bruise can become 7 cm. Low-confidence pages need a human read.
• Hallucination. Generative AI can state a fact no page supports, and notes drafted by ambient AI scribes can carry errors into the chart itself.

The fixes are plain: a page-level citation on every line and a human-in-the-loop reviewer who reads the cited page before anything is used. Lawyers were sanctioned under Rule 11 for fabricated AI-generated citations in Mata v. Avianca (S.D.N.Y. 2023); a fabricated clinical fact in an expert disclosure is the same failure.

Vendor checklist for HIPAA compliant AI on a child abuse file

Legal AI tools also help with deposition summaries and redacting the child's identifiers in exhibits, again as cited drafts a person checks.

Adjust the legal basis to your role and any protective order.

RE: [CHILD INITIALS], DOB [DATE], MRN(s) [IF KNOWN]
Date range: [BIRTH OR START] to [END]
Legal basis: [SUBPOENA / COURT ORDER / AGENCY AUTHORITY]

Please produce for the date range:
[ ] All encounter notes: well-child, sick, urgent care,
    telehealth, ED, inpatient, with triage and nursing notes
[ ] Growth charts and developmental milestones
[ ] Portal messages with attachments and photographs
[ ] Nurse advice line records
[ ] All imaging in native DICOM with reports, including
    initial and follow-up skeletal surveys
[ ] Ophthalmology notes and retinal photographs
[ ] Hematology, genetics, endocrinology consults and results,
    including send-out results received after discharge
[ ] Child protection team notes and clinical photographs
[ ] Social work notes and any agency report record
[ ] Birth records if this facility delivered the child

If any category is withheld or does not exist, please say
so in writing, with the reason.
Contact: [NAME, ROLE, PHONE, EMAIL]

CHILD: [INITIALS]  DOB: [DATE]  REVIEWER: [NAME]

FINDING: [e.g., cheek bruise / subdural hemorrhage]
  First documented:  [DATE, TIME] by [ROLE], p. [__]
  Source words:      "[QUOTE, WITH ANY HEDGE]"
  Age / mobility:    [__ weeks] / [as charted], p. [__]
  Histories:         [DATE, TELLER, QUOTE, p. __] (repeat)
  Imaging:           [STUDY, DATE, IMAGE COUNT, READER, p. __]
  Follow-up imaging: [DONE p. __ / NOT DONE, REASON p. __]
  Alternatives documented:
    Bleeding:     [TEST OR CONSULT, p. __ / NONE FOUND]
    Bone:         [TEST OR CONSULT, p. __ / NONE FOUND]
    Metabolic:    [TEST OR CONSULT, p. __ / NONE FOUND]
    Accidental:   [HOW TESTED, p. __ / NONE FOUND]
    Birth:        [RECORD, p. __ / NOT REQUESTED]
  Gaps:              [MISSING RECORD, IMPLIED BY p. __]
  For expert:        [QUESTIONS ONLY]

DATE | AGE (WEEKS) | SOURCE | MOBILITY | SKIN EXAM | MOUTH EXAM
     | HISTORY (TELLER) | PLAN | CITE

Rules:
- Compute age from DOB; never copy a rounded age.
- Write "not documented" for any region not mentioned.
  Never write "normal" for silence.
- PLAN: survey, consult, report, photo, or "none".

Flag for expert review when:
[ ] a bruise or mouth injury appears before cruising
[ ] the child is 4.99 months or younger with any bruise
[ ] a record is referenced but not produced

Medrecords AI is medical chronology software. It works on the records you upload; it does not request or retrieve records from providers or agencies. On a child abuse file it does 5 things:

• Builds a cited chronology of every encounter across every source, with a citation on every line, so the chapter 9 timeline starts as a draft.
• Flags missing records the file implies, such as a portal reply that mentions an unproduced photo, each flag cited to the line that implies it.
• Flags pages that appear to belong to another patient, such as a sibling's note in the index child's chart.
• Routes scanned and handwritten pages through OCR and flags low-confidence pages for a human read.
• Loads CT, MRI and survey studies as DICOM in a browser viewer through imaging review, so your radiology expert can cite an image like a page.

What it does not do: read images for findings, date fractures, judge whether a history fits an injury, score the likelihood of abuse or give legal advice. Flags are signals, not verdicts. It runs under SOC 2 and HIPAA with a signed BAA (see security and HIPAA). Self-Service bills 10 cents a deduplicated page, down to 5 cents at volume, duplicates free; Enterprise On-Prem is an annual license. Related: child custody evaluation record review and record review for legal nurse consultants.

What is a sentinel injury in child abuse?

A minor injury, usually a bruise or mouth injury, found in an infant before a more serious abusive injury, suspicious because the baby could not cruise or the explanation was implausible. In a 2013 study, 27.5% of 200 definitely abused infants had 1; none of 101 nonabused controls did.

What does TEN-4-FACESp stand for?

Bruising on the torso, ear or neck; on the frenulum, angle of the jaw, cheeks, eyelids or subconjunctivae; any bruise at 4.99 months or younger; or patterned bruising. It was 95.6% sensitive and 87.1% specific in children under 4. A positive result calls for more evaluation, not a diagnosis.

When is a skeletal survey repeated?

The ACR says a follow-up survey may be indicated; a 2017 review puts it at 10 to 14 days. In 1 large study, 15.6% of follow-up surveys found a new fracture and 6.9% were reassuring.

Is abusive head trauma the same as shaken baby syndrome?

The AAP replaced "shaken baby syndrome" with "abusive head trauma" in 2009 because injury can come from shaking, impact or both. The diagnosis is contested in court: a 2018 consensus statement supports it, and a 2017 systematic review found insufficient evidence on the triad's accuracy for shaking.

What medical conditions can look like child abuse?

AAP reports discuss bleeding disorders, bone fragility such as osteogenesis imperfecta, and metabolic bone disease, alongside accidental and birth-related injury. The records show which were tested.

Does HIPAA allow a hospital to report suspected child abuse?

Yes. 45 CFR 164.512(b)(1)(ii) permits disclosure to a government authority authorized to receive reports of child abuse or neglect. Who must report is state law.

Can AI tell whether a child was abused?

No. AI software can build a cited timeline, compute age at every visit, pull every stated history and flag missing records. Whether injuries were inflicted is for experts and the court.

Can ChatGPT summarize a child abuse medical record?

A general chatbot gives no page citations, can drop hedges like "possible," and can invent facts, and a minor's records need a HIPAA compliant tool with a BAA. Use software that cites every line, and check each cite.

How does automated review handle photographs and x-rays?

It should index them on the timeline, not interpret them. That is expert work.

Study figures were checked against PubMed abstracts, the ACR and SPR parameter against its full text, and statutes and opinions on Cornell and CourtListener, in September 2026. Quotes are verbatim, except that dashes are rendered as colons or "to." The exam note, timeline, gap ledger, names and page numbers in the examples are hypothetical.

• Pierce and colleagues, JAMA Netw Open 2021 and Pediatrics 2010: TEN-4-FACESp and TEN-4.
• Sheets and colleagues, Pediatrics 2013: sentinel injuries.
• Sugar and colleagues, Arch Pediatr Adolesc Med 1999: bruising by mobility.
• Jenny and colleagues, JAMA 1999: missed abusive head trauma.
• Harper and colleagues, Pediatrics 2013 and Lindberg and colleagues, Pediatrics 2012: follow-up surveys and contacts.
• ACR and SPR skeletal survey practice parameter (revised 2021, amended 2023) and Aertsen, J Belg Soc Radiol 2017.
• AAP: AHT 2009, AHT 2020, imaging 2009, fractures 2014, evaluation 2015, bleeding disorders 2022.
• Choudhary and colleagues, Pediatr Radiol 2018 and Lynøe and colleagues, Acta Paediatr 2017.
• Kempe and colleagues, JAMA 1962; Block and Palusci, J Pediatr 2006; AMA reporting, 2010; Council of Pediatric Subspecialties.
• 42 USC 5106a; 42 USC 5101 notes; 45 CFR 164.512; 45 CFR 164.502; FRE 702.
• Cavazos v. Smith, 565 U.S. 1 (2011); Landeros v. Flood, 17 Cal. 3d 399 (1976); Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023).

Related guides: how to read a radiology report, TBI imaging and diffuse axonal injury, blunt force trauma medical records and how to read an autopsy report.