Medical report analysis
Phase 2: a consistent injury breakdown for every Official Injury Claim medical report
Background
Since May 2021, road traffic injury claims in England and Wales worth up to £5,000 for the injury (and up to £10,000 in total) go through the Official Injury Claim portal. Every claim needs a medical report from an expert accredited through MedCo, the body that manages the supply of these reports. The report sets out each injury and how long it is expected to last.
Whiplash injuries lasting up to two years are valued from a fixed tariff in the Whiplash Injury Regulations 2021, with a small uplift where there is also a minor psychological injury. Following the 2025 amendment, for accidents on or after 31 May 2025, the tariff runs from £275 to £4,830. Any other injury, such as a knee, a wrist or a seatbelt bruise to the chest, is valued in the ordinary way. Claims with both kinds are "mixed injury" claims. Following the Supreme Court's decision in Hassam v Rabot (2024), each part is valued separately and a deduction made for overlap, with the total never less than the non-tariff part alone would have received.
This is why the split matters. Published commentary has put the share of portal claims with a non-tariff element at close to half (we have not verified this figure against the Ministry of Justice's own statistics, and One Claim's own mix may differ). Every one of those needs a handler to value the non-tariff part by judgement rather than from a table, and each carries more cost and more room for disagreement.
What is the count for? For example: management information for insurer clients, setting reserves, spotting claimant representatives whose reports show unusual patterns, or checking offers before they go out. The answer changes what "cross-reference" needs to mean.
What we extract from each report
Reading a medical report is the one step here that needs inference: layouts differ between agencies and experts, and the facts are written in prose. Claude's job is limited to filling in a fixed record from the report. Every item links to the page and passage it came from, so a handler can check it in seconds. Everything after extraction (classification, tariff amounts, cross-checks, counts) is done by rules.
Where a medical agency uses a consistent layout, fields such as the MedCo reference, dates and the prognosis table are read by position-based rules instead, and Claude is only asked for what those rules could not find.
| Group | Fields |
|---|---|
| Report | Expert name and discipline, MedCo reference, date of examination, date of report, whether it is an initial or addendum report, whether records were reviewed |
| Claimant and accident | Name, date of birth, date of accident, vehicle position, account of the accident as recorded by the expert, time between accident and examination |
| Each injury | Body area, description, onset, current severity, prognosis in months, whether it is resolved, whether it is an exacerbation of an earlier condition |
| Psychological | Symptoms described, whether the expert calls it minor, prognosis |
| Treatment | Treatment so far, recommended treatment (for example, a number of physiotherapy sessions), further reports or referrals recommended |
| Other | Time off work, effect on daily life, previous accidents or claims mentioned, anything the expert flags about consistency |
Tariff or non-tariff
Each injury is then put in one of these groups by rule, not by the model. Claude reports the facts for each injury (body area, tissue type, prognosis in months, whether it is psychological); a fixed decision table maps those facts to a group. The table follows the definition in the Civil Liability Act 2018, which treats whiplash as a soft tissue injury (a sprain, strain, tear, rupture or lesser damage of a muscle, tendon or ligament) in the neck, back or shoulder, or an injury associated with one, arising from a road traffic accident while in a vehicle, and lasting no more than two years.
| Group | Meaning | Example |
|---|---|---|
| Tariff whiplash | Soft tissue injury to neck, back or shoulder, prognosis up to 24 months | Neck strain, 6 months |
| Tariff psychological | Minor psychological injury alongside tariff whiplash | Travel anxiety, 4 months |
| Non-tariff | Any other injury | Knee contusion, wrist sprain, chest bruising from a seatbelt |
| Outside tariff by duration | Neck, back or shoulder soft tissue injury expected to last over 24 months | Lower back, 30 months |
| Needs a person | The extracted facts do not fit the table: unclear, contradictory, or a borderline body area | "Upper back and chest wall pain" |
Each report then gets an overall label: whiplash only, mixed, or non-tariff only. The tariff amount for the whiplash part is looked up from a table held in the software (not recalled by the model), keyed on accident date and longest whiplash prognosis. Non-tariff injuries are listed for the handler to value; Postbot does not value them.
We will load the full tariff tables from the 2021 Regulations and the 2025 amendment (statutory instrument 2025/615) and have them checked by a One Claim handler before use.
Cross-referencing
The customer notes say the reports need cross-referencing "to things". These are the checks we think would be most useful. Each produces a flag with a reason, not a conclusion. Once the facts are extracted, almost all of them are rules comparing one field with another.
| Check | How | What it compares |
|---|---|---|
| Report against claim form | rules | Injuries and durations claimed on the portal against those in the report, field by field |
| Timing | rules | Days from accident to examination, and from examination to report, against thresholds One Claim sets |
| Treatment against prognosis | rules | Recommended physiotherapy sessions against prognosis length, using a threshold table |
| Report against vehicle damage | rules | Injury prognosis against the engineer's repair cost or damage category, to highlight low-damage, high-injury combinations |
| Earlier history | rules | Whether previous accidents were recorded as facts in extraction, against prior claims held by One Claim |
| Patterns across reports | rules | Share of mixed and non-tariff claims and typical prognosis, by representative, agency and expert. A database query, not a judgement on any one claim. |
| Report against incident account | inference | The expert's prose account of the accident against the driver's prose account. This is the one check that compares two pieces of free text, so it needs Claude. It reports specific differences with quotes from both; a rule decides whether they raise a flag. |
These flags are prompts for a trained handler. The system will not state that a claim is fraudulent or exaggerated, and outputs should not be used to make decisions about a claimant without human review.
Outputs
For each claim
- One-page summary: injuries, groups, overall label, tariff amount for the whiplash part
- Each fact linked to its page in the report
- Cross-reference flags with reasons
- Handler can correct any field; corrections are kept
Across the portfolio
- Counts: whiplash only, mixed, non-tariff only, needs a person
- Breakdown by month, client, representative, medical agency
- Distribution of prognosis periods
- Export to spreadsheet for reporting to insurer clients
Getting the reports
The Official Injury Claim portal lets compensators work through a website or connect their own systems (the service calls this an application-to-application link). The technical specification for the link is only given to registered organisations, so we have not seen it. Three routes, in order of preference:
- Existing link: if SiteView already connects to the portal, the reports may already be stored against claims in SiteView, and we read them from there.
- Direct link: One Claim, as a registered compensator, obtains the specification and we build a read-only connection for medical reports.
- Manual export: staff download reports into a watched folder. Slower, but it works on day one and is how we would process any backlog.
How does One Claim use the portal today, and does SiteView connect to it? Roughly how many reports are in the backlog, and how many arrive each month? We also need to check the portal's terms of use for compensators allow automated processing of documents taken from it.
Backlogs suit Anthropic's batch service, which processes large volumes within 24 hours at half the normal price. New reports are processed as they arrive.
Limits
- Medical reports are health data, a special category under UK data protection law. This work needs a data protection impact assessment before any real report is processed. See data protection.
- Report formats vary between agencies. Accuracy must be measured on a sample of One Claim's real reports, checked by a handler, before the counts are relied on.
- Handwritten or poor-quality scans may be read less reliably. These will be flagged rather than guessed.
- The system classifies and flags. Valuation of non-tariff injuries and any decision on the claim stay with the handler.