Worked example: 1 synthetic 40-page record, summarized 3 ways
Answer in brief: we built a synthetic 40-page workers' comp low back file and summarized it by hand, with the IME template, and as an AI-style draft with page citations. The template version caught all 5 issues we planted. The hand narrative caught 2. The AI-style draft caught 1 and carried 3 errors, all of which verification against the cited pages caught.
Synthetic, no real patient. Every name, date and finding on this page is invented for teaching. We wrote all 3 summaries ourselves; the AI-style draft shows the error types published studies report and is not output from any product.
Limits: We have not timed the listed tools on a shared test file: time figures come from the cited studies and vendor pages, and the worked example is synthetic.
The file: 40 pages as received
Patient R.T. (synthetic), 52, warehouse worker. Lifting injury at work on 2025-03-04. The carrier asks for a records review before an IME, with 5 referral questions on page 39. Pages arrive in the order the carrier assembled them, not by date.
| Pages | Date | Document | Key content |
|---|---|---|---|
| 1-3 | 2025-03-04 | ED visit, City General (synthetic) | Lifted a 40 lb box; low back pain to left buttock. SLR negative both sides. X-ray: no fracture, mild L5-S1 disc space narrowing. Lumbar strain; cyclobenzaprine, ibuprofen; off work 3 days |
| 4-5 | 2025-03-04 | Same ED note, faxed by the PCP | Exact duplicate of pages 1-2 |
| 6-9 | 2025-03-07 | Occupational medicine, initial | Pain 7/10, radiating down left leg to calf. SLR positive left at 40 degrees. Decreased sensation left lateral foot. Lumbar strain with left radicular symptoms. PT 2 times a week for 6 weeks; modified duty, lift up to 10 lb |
| 10-11 | 2025-03-21 | Occupational medicine, follow-up | Worse; numbness left lateral foot. MRI ordered |
| 12-13 | 2025-03-28 | MRI lumbar spine | "L5-S1 left paracentral disc extrusion contacting the traversing left S1 nerve root." L4-5 mild bulge, no stenosis. Mild facet arthropathy L4-S1 |
| 14-19 | 2025-04-02 to 2025-04-30 | Physical therapy, 6 visits | Pain 7/10 falling to 5/10. Page 16 (2025-04-16) reads "right leg radicular pain"; every other note says left |
| 20-21 | 2025-05-12 | Pain management | Left L5-S1 transforaminal epidural injection recommended; gabapentin 300 mg at night started |
| 22 | 2025-05-19 | Procedure note | Left L5-S1 transforaminal epidural steroid injection, no complications |
| 23-24 | 2023-08-15 | Primary care note (before the injury) | "Low back pain, intermittent, 2 years, no radiation." Naproxen 500 mg as needed. No imaging |
| 25 | 2023-02 to 2025-06 | Pharmacy printout | Naproxen fills 2023-02, 2023-08, 2024-01; cyclobenzaprine 2025-03; gabapentin 2025-05 |
| 26-28 | 2025-06-16 | Spine surgery consult | Pain 6/10. Left SLR positive at 45 degrees; left plantarflexion 4/5; left ankle reflex decreased. Microdiscectomy discussed; patient chooses to continue conservative care |
| 29-30 | 2025-07-01 | Occupational medicine | Better after injection, 4/10. Modified duty continues. "EMG/NCS left lower extremity ordered" |
| 31 | 2025-03-07 to 2025-07-01 | Work status forms | Modified duty, 10 lb limit, throughout |
| 32-33 | 2025-08-12 | Occupational medicine | Pain 3/10. SLR negative both sides; strength 5/5. Trial of full duty from 2025-08-18 |
| 34 | 2025-08-05 | PT discharge | 18 visits in total; most goals met |
| 35-36 | 2025-03 to 2025-08 | Billing ledger | Charges by provider; no clinical content |
| 37 | 2025-03-28 | MRI report again, from the surgeon's packet | Exact duplicate of page 12 |
| 38 | 2025-09 | Records request cover letters | No clinical content |
| 39 | 2025-09-10 | Carrier referral letter | 5 questions: diagnosis; causation and any apportionment; whether treatment was reasonable and necessary; MMI; work capacity |
| 40 | 2025-03-05 | Employer's first report of injury | Lifting a box, low back; reported the next day |
What we planted: 3 duplicate pages (4, 5, 37), a prior history of the same body part split across 2 documents (pages 23-25), a laterality conflict (page 16), an EMG ordered with no report in the file (page 29), and an ED exam with a negative SLR 3 days before a positive test. A good summary should handle all 5.
Reading time at published paces: 40 pages at 40 to 100 pages an hour is 24 to 60 minutes before any writing (benchmarks and sources on the manual vs AI page). We did not time ourselves; these summaries are teaching examples, not a speed test.
Method 1: by hand, as a narrative
A reviewer reads front to back and writes prose as they go, adding page numbers for the main facts. This is how most hand summaries look.
R.T. is a 52-year-old warehouse worker who injured his low back lifting a 40 lb box at work on 2025-03-04 (p. 1). The ED diagnosed lumbar strain; x-rays showed no fracture (p. 2). Occupational medicine on 2025-03-07 found left leg radiation, a positive left straight leg raise and decreased sensation in the left lateral foot, and placed him on modified duty (p. 6-9). Symptoms worsened and an MRI on 2025-03-28 showed an L5-S1 left paracentral disc extrusion contacting the left S1 nerve root (p. 12).
He completed physical therapy with gradual improvement (p. 14-19, 34), received a left L5-S1 transforaminal epidural injection on 2025-05-19 (p. 22), and saw a spine surgeon on 2025-06-16 who found left S1 deficits and discussed microdiscectomy; he chose conservative care (p. 26-28). By 2025-08-12 pain was 3/10 with a normal exam, and he began a trial of full duty on 2025-08-18 (p. 32-33).
A 2023 primary care note documents intermittent low back pain for 2 years treated with naproxen, without radiation (p. 23). Pages 4-5 and 37 repeat earlier records.
What it caught: the clinical story, the prior history, the duplicates and the main page numbers. What it missed: the EMG that was ordered and never appears (page 29), the laterality conflict on page 16, which the reader smoothed over, and the change from a negative SLR in the ED to a positive test 3 days later. It also doesn't answer the questions in order. A reader who wants to check "gradual improvement" has 7 pages to search.
Method 2: the template and extraction grid
Same reading, but every encounter goes into the extraction grid first, and the summary follows the IME template. Here is the grid, shortened to the rows that matter for the questions.
| Date | Page | Provider | Findings (quoted or close) | Diagnosis | Plan or status | Flag |
|---|---|---|---|---|---|---|
| 2023-08-15 | 23-24 | PCP | "Low back pain, intermittent, 2 years, no radiation" | Low back pain | Naproxen as needed | Prior history, same body part |
| 2023-02 to 2024-01 | 25 | Pharmacy | Naproxen filled 3 times | Prior history | ||
| 2025-03-04 | 1-3 | ED | Pain to left buttock; SLR negative both sides | Lumbar strain | Off work 3 days | Dup at 4-5 |
| 2025-03-07 | 6-9 | Occ med | SLR positive left 40 degrees; decreased sensation left lateral foot | Strain with left radicular symptoms | PT; modified duty 10 lb | |
| 2025-03-28 | 12-13 | Radiology | "L5-S1 left paracentral disc extrusion contacting the traversing left S1 nerve root" | Dup at 37 | ||
| 2025-04-16 | 16 | PT | "Right leg radicular pain" | Conflict: all other notes left | ||
| 2025-05-19 | 22 | Pain mgmt | Left L5-S1 TFESI | |||
| 2025-06-16 | 26-28 | Spine surgery | Left SLR positive 45 degrees; plantarflexion 4/5 left; left ankle reflex decreased | Left S1 radiculopathy | Conservative care by choice | |
| 2025-07-01 | 29-30 | Occ med | 4/10 after injection | EMG/NCS ordered; modified duty | EMG report not in file | |
| 2025-08-12 | 32-33 | Occ med | 3/10; SLR negative both sides; strength 5/5 | Full duty trial from 2025-08-18 |
The summary written from that grid:
Records reviewed: 40 pages from 7 sources; 3 exact duplicates removed (p. 4-5 copy p. 1-2; p. 37 copies p. 12). Not received: EMG/NCS ordered 2025-07-01 (p. 29).
Summary for Q1 to Q5: lifting injury at work 2025-03-04 (p. 1, 40), with left leg radicular findings from 2025-03-07 (p. 6-9) and an MRI showing an L5-S1 left paracentral extrusion contacting the left S1 root (p. 12). Treated with PT (18 visits, p. 34), 1 left L5-S1 epidural injection (p. 22) and medication. The surgeon documented left S1 deficits and discussed microdiscectomy; the patient chose conservative care (p. 26-28). On 2025-08-12: pain 3/10, SLR negative, strength 5/5, trial of full duty from 2025-08-18 (p. 32-33). No later records received.
Pre-injury history: intermittent low back pain for 2 years, no radiation, on naproxen as needed (p. 23-24); naproxen filled 3 times 2023-02 to 2024-01 (p. 25). No prior imaging or leg symptoms found.
Flags: (1) ED SLR negative on 2025-03-04 (p. 2), positive on 2025-03-07 (p. 6). (2) PT note p. 16 says right leg; all other records say left. Likely a documentation error; noted, not relied on. (3) EMG/NCS ordered (p. 29), no report received; request it before the exam.
What it caught: all 5 planted issues. The cost is the grid itself: about 1 row per encounter, which is the extraction time the BMC study measured at roughly a minute per report.
Method 3: AI-assisted, then verified
Software reads the file and returns a draft with a page cite on each line. Below is an illustrative draft we wrote with the 3 kinds of error the published studies report most: an omission, an inaccuracy and a hallucination. It is not output from any product.
Work-related lifting injury on 2025-03-04 (p. 1). ED: lumbar strain, x-ray negative for fracture (p. 2). Occupational medicine: left radicular symptoms, positive left SLR, modified duty (p. 6-9). MRI 2025-03-28: L5-S1 paracentral disc extrusion contacting the right S1 nerve root (p. 12). PT 18 visits with improvement (p. 14-19, 34). Left L5-S1 TFESI 2025-05-19 (p. 22). Spine surgery consult: left S1 deficits, conservative care chosen (p. 26-28). EMG/NCS confirmed left S1 radiculopathy (p. 29). Full duty trial from 2025-08-18 (p. 32-33). No prior history of low back complaints documented. Duplicates removed: p. 4-5, 37.
Verification takes the draft line by line against its cited pages, then searches for what is missing:
| Draft line | Error type | How verification caught it | Correction |
|---|---|---|---|
| MRI: contacting the right S1 root (p. 12) | Inaccuracy (laterality) | Opened p. 12: report says left. The draft likely picked up the PT note on p. 16 | Left S1 root; flag p. 16 as a conflict |
| EMG/NCS confirmed left S1 radiculopathy (p. 29) | Hallucination | Opened p. 29: the test is ordered, no result. No EMG report anywhere in the index | EMG ordered, report not received; request it |
| No prior history documented | Omission stated as a finding | Searched the index for dates before 2025-03-04: p. 23-25 exist | Add the 2023 PCP note and pharmacy fills |
The other cited lines checked out against their pages. After correction, the content matches the template summary above, with a page cite on every line. The value of the cites showed in the check: each error took 1 page to disprove, instead of a reread of the file.
The hallucinated EMG is the most dangerous of the 3: it supports causation, it carries a real page number, and the page does mention an EMG. Only opening the page shows that the test was ordered, never reported. That is why the quality checks say to open every cited page behind a line your opinion rests on, not only lines that look wrong.
What each method caught
| Planted issue | By hand | Template | AI draft, unverified | AI draft, verified |
|---|---|---|---|---|
| 3 duplicate pages identified | Yes | Yes | Yes | Yes |
| Prior low back history (p. 23-25) | Yes | Yes | No | Yes |
| Laterality conflict (p. 16) | No | Yes | No, and carried into the MRI line | Yes |
| EMG ordered, no report (p. 29) | No | Yes | No, and invented a result | Yes |
| ED SLR negative, then positive | No | Yes | No | Yes, on the search for gaps |
| Page cite on every fact | Main facts only | Yes | Yes | Yes |
This table describes the 3 versions we wrote, not a measured comparison of methods. What it does show is where each method's risk sits: the hand narrative loses what the reader didn't think to write down, and the AI draft states things with confidence that the page doesn't support. The template is the check that works for both.
Questions about the worked example
Is this a real patient's medical record?
No. The file, the patient R.T., the providers and every finding are synthetic, written for teaching. No real patient's records were used.
Did a real AI tool write the AI draft?
No. We wrote it to show the 3 error types reported most in published studies of AI summaries: omissions, inaccuracies and hallucinations. It is not output from any product, ours included.
Why did the template catch more than the hand narrative?
The grid forces a row for every document and a flag column, so a missing EMG report or a side conflict has a place to land. In a narrative, the reader only writes what they notice at the moment.
How would I use this example?
As a calibration file: give it to a new reviewer or a tool you are testing, and see which of the 5 issues come back. Any summary that misses the hallucination check on page 29 is not ready to sign.