Skip to main content

ReadYourLab Research · Report 1 · July 2026

The Patient Second Read

What 1,815 North American imaging studies reveal about how patients use AI to understand their CT and MRI scans — and how readable the radiology reports they carry really are.

Since the 21st Century Cures Act, imaging reports in the United States are released to patient portals immediately — often before any clinician has discussed them. What do patients do in that gap? Survey studies have asked them. This report observes them: it is drawn from anonymized, aggregate usage data of ReadYourLab, a service where patients voluntarily upload their own imaging (original DICOM files) and radiology reports for an AI explanation. To our knowledge, it is the first published behavioral dataset of the "patient second read" — measured, not surveyed.

Published July 22, 2026 Dataset: Apr 21 – Jul 21, 2026 Geography: United States & Canada

Key findings

  • 1.The second read happens fast. 44% of North American patients uploaded their scan for an AI explanation within 3 days of the scan being performed; 13% did so the same day. The median gap was 6 days.
  • 2.A third of second reads concern imaging the radiologist finished long ago. 31.5% of studies were uploaded more than 30 days after the scan — these patients already have a signed report, and are looking for a better explanation of it.
  • 3.Nearly 1 in 5 second reads happens at night. 18.5% of uploads occurred between 10 p.m. and 6 a.m. local time. Friday is the busiest day of the week.
  • 4.The report is not the end of the conversation. 6 in 10 paying patients asked at least one follow-up question after receiving an AI explanation — averaging 5.4 questions each.
  • 5.The reports patients carry are not written for them. 97% of English-language radiology reports in our corpus score above the 8th-grade reading level recommended for patient materials (median: grade 11.5). The median report contains 14 terms that require a medical dictionary. One in five is a paper scan with no machine-readable text at all.
44%
seek an AI explanation within 3 days of their scan
1 in 5
uploads happen between 10 p.m. and 6 a.m.
97%
of radiology reports exceed the recommended patient reading level
6 in 10
patients ask follow-up questions after the AI explanation

1 · The 72-hour window

For every study, the DICOM file records the date the scan was performed. Comparing it with the moment the patient uploaded that scan measures how long patients sit with imaging before seeking an explanation on their own. The answer: not long. Half of all North American second reads happen within 6 days of the scan; 44% within 3 days. In the immediate-release era, the race between the patient's portal and the patient's physician appointment is usually won by the portal — and the patient does not wait.

The right side of the chart tells a second story. 31.5% of studies were uploaded more than 30 days after the scan. By then, the signed radiologist report has long existed. These patients are not impatient; they are unsatisfied. They return to old imaging looking for an explanation they never felt they got.

Time from scan to patient upload Share of unique studies, US & Canada, Apr 21–Jul 21 2026 (n = 1,808)
0 10% 20% 30% 12.9% 31.2% 9.8% 14.5% 13.7% 5.8% 12.0% Same day 1–3 days 4–7 days 8–30 days 1–6 months 6–12 months > 1 year 44% within 3 days 31.5% more than 30 days after the scan
Data table
Time from scan to uploadShare of studies
Same day12.9%
1–3 days31.2%
4–7 days9.8%
8–30 days14.5%
31–180 days13.7%
181–365 days5.8%
More than 1 year12.0%
How this was measured. Each uploaded DICOM file carries a StudyDate tag written by the scanner at acquisition time. For each unique study (deduplicated by Study Instance UID and user), we computed the number of days between StudyDate and the upload timestamp. 1,815 unique CT/MRI studies from US/Canada-located users were uploaded in the window; 1,808 had a parseable StudyDate. For comparison, the median for users outside North America was 3 days — the fast second read is a global behavior, not an American one.

2 · The after-hours reality

Portals do not have office hours, and neither does scan anxiety. 18.5% of North American uploads occurred between 10 p.m. and 6 a.m. local time — nearly one in five patients sat down with their imaging when no clinic was open and no physician was reachable. Activity stays elevated through the late evening, with a visible bump around midnight.

Uploads by hour of day Unique studies, US & Canada, local time approximated as US Central (n = 1,815)
12am 3am 6am 9am 12pm 3pm 6pm 9pm Night (10pm–6am): 18.5% of uploads
Data table
Hour (local)UploadsHour (local)Uploads
12am4612pm107
1am501pm132
2am252pm125
3am233pm135
4am254pm122
5am165pm112
6am216pm131
7am247pm111
8am398pm99
9am699pm79
10am7710pm68
11am9711pm82
Uploads by day of week Unique studies, US & Canada (n = 1,815)
12.6% 14.5% 14.8% 15.4% 18.1% 13.7% 10.9% Mon Tue Wed Thu Fri Sat Sun
How this was measured. Upload timestamps (UTC) were shifted to US Central Time as a single approximation for the US/Canada cohort; the true spread of North American time zones adds roughly ±2 hours of noise, which the wide 10 p.m.–6 a.m. night window absorbs. The Friday peak (18.1% of weekly volume) is consistent with reports being finalized and released to portals before the weekend — patients receive results on Friday and face two days without anyone to call. For comparison, the night share among European users was 15.6%.

3 · Who seeks a second read

The AI-era imaging patient is not who you might expect. More than half of adult studies (52%) belong to patients in their 30s and 40s — a working-age population, slightly more male than female (54.7% vs 45.3%). MRI dominates CT by more than 3 to 1, and musculoskeletal imaging — joints, extremities and the spine — accounts for just over half of all studies. This is the demographic of back pain, sports injuries and long waits for a specialist appointment: patients with a scan in hand, a portal login, and questions.

Patient age at time of scan Share of adult studies, 18+ (n = 1,723 with known age)
28.3% 23.7% 18–29 30s 40s 50s 60s 70s 80+ 52% of studies
Modality & sex Modality: all studies (n = 1,814 CT/MRI). Sex: adult studies with sex recorded (n = 1,708)
MRI 76.9%CT 23.1%

A further 259 single-image X-ray analyses were run in the same window (counted separately from DICOM studies).

Male 54.7%Female 45.3%

Sex as recorded in the DICOM header by the imaging facility.

What gets a second read: anatomy Share of unique studies, classified from DICOM study descriptions (n = 1,815)
Joints & extremities 26.8% Spine 23.9% Head & brain 17.9% Abdomen & pelvis 12.1% Chest & cardiac 5.6% Other / vascular 2.8% Unclassified 11.0%

Musculoskeletal imaging (joints, extremities and spine) together: 50.7% of all studies.

How this was measured. Age is computed from the patient birth year and scan year recorded in the DICOM header; sex comes from the same header, as entered by the imaging facility. Demographic charts cover adult patients (18+) only by policy. Anatomy was classified from DICOM study and series descriptions using a multilingual keyword ruleset into six generic regions; 11% of descriptions (scanner protocol shorthand like “t2_tra”) could not be classified and are shown as such rather than redistributed.

4 · The report is not the end of the conversation

An explanation, it turns out, is where patient engagement starts, not where it ends. Among paying North American users — the cohort with unrestricted access to follow-up chat — 60.5% asked at least one question after receiving their AI explanation, averaging 5.4 questions each. These are the questions that historically had nowhere to go: the radiologist is out of reach, the referring physician's appointment is weeks away, and the report has already been read — at 11 p.m., on a Friday.

Engagement is heavily skewed. Half of asking patients ask one to three questions (median: 3) — but one in ten asks thirteen or more, and nearly one in five asks more than ten. For a minority of patients, the question load their imaging generates is enormous, and nothing in the current care pathway is sized to absorb it.

A second signal points the same direction. We added the option to upload the radiologist’s written report together with the scan. With no promotion, 6% of all studies now arrive with the signed report attached — patients explicitly requesting a comparison between what their radiologist wrote and what the AI sees, rather than a first read.

60.5%
of paying patients ask ≥1 follow-up question
5.4
average questions per asking patient
6%
of studies arrive with the radiologist’s report attached for comparison
How many questions patients ask Distribution among paying patients who asked at least one question, US & Canada (n = 291)
25.1% 31.6% 17.9% 12.7% 7.6% 5.2% 1 2–3 4–5 6–10 11–20 20+ Follow-up questions asked 18% ask more than 10
Data table
Questions askedShare of asking patients
125.1%
2–331.6%
4–517.9%
6–1012.7%
11–207.6%
More than 205.2%
How this was measured. Follow-up questions are user-written messages inside scan-analysis conversations; the denominator is North American users with a completed purchase in the study window (n = 481), since follow-up chat is a paid feature — free-tier engagement cannot be measured and is excluded rather than estimated. Among the 291 patients who asked, the distribution is right-skewed: median 3 questions, 90th percentile 13 — the 5.4 average is pulled up by a highly engaged tail. The report-attachment rate counts unique studies uploaded after the feature’s June 3 launch that included a report document (5.9% overall; the North American subset matches at 5.9%, n = 66).

5 · The reports patients carry

Alongside their scans, patients upload the radiologist’s written report — a corpus of 259 real-world report documents, exactly as patients received them. Health-literacy guidance (AMA, NIH) recommends patient-facing material be written at or below an 8th-grade reading level. 97% of the English-language reports in our corpus exceed it. The median report scores at grade 11.5 on the Flesch-Kincaid scale; 39% score at college level or above.

Readability formulas actually understate the problem. The median report contains 14 distinct terms that warrant a medical dictionary (“spondylolisthesis,” “attenuation,” “enhancement”), and 80% contain ten or more. One in three uses hedge phrasing (“cannot exclude,” “clinical correlation recommended”) whose purpose is opaque to lay readers. One in ten cites a formal grading system — Modic, Bosniak, Fazekas — with no explanation attached. And 19% of all report documents are scans or photos of paper with no machine-readable text: those patients cannot copy a term into a search engine, a translator, or an accessibility tool. Visual review of that scanned fifth found that more than a third are phone photographs of a paper document, and nearly one in five carries handwritten annotations — in 2026, a meaningful share of patients still receives their imaging results as a piece of paper.

Reading level of radiology reports Flesch-Kincaid grade of English-language, text-extractable reports (n = 89)
Recommended for patients: grade 8 or below 97% of reports are above it 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20+ Flesch-Kincaid grade level median 11.5
14
medical-dictionary terms in the median report
19%
of report documents are paper scans with no selectable text
34%
contain hedge phrasing (“cannot exclude…”)
1 in 10
cites a formal grading system (Modic, Bosniak…) unexplained
How this was measured. The corpus comprises 259 report documents voluntarily uploaded by patients alongside their imaging (56% of the English subset from North American users; 14 languages in total). 209 documents had extractable text; 50 (19%) were image-only scans or photos of paper. For the 89 English text documents we computed Flesch-Kincaid grade level, counted distinct matches against a 1,166-entry radiology glossary (our public medical dictionary), and flagged 14 hedge phrases and 10 named grading systems. The 50 image-only documents were individually reviewed (AI-assisted visual classification): 44 are radiologist reports, 18 are phone photos of paper, and 9 contain handwriting. Reports are real-world documents as patients received them — not a curated hospital sample. No report content is quoted or published; all statistics are aggregate.

Methodology & data ethics

Dataset. Anonymized, aggregate usage data of ReadYourLab between April 21 and July 21, 2026. The North American cohort covers users located in the United States and Canada (determined by IP geolocation at signup) and comprises 1,815 unique CT/MRI studies — deduplicated by DICOM Study Instance UID, so one study equals one patient imaging event — plus 259 patient-uploaded report documents (all languages). Administrative and internal test accounts were excluded.

Self-selection, stated plainly. This is not a random sample of imaging patients; it is the population that chose to seek an AI explanation of their imaging. That is precisely the phenomenon under study — the numbers describe the patients who take this step, not all patients. Where relevant we report the non-North-American comparison, which tells the same story (median scan-to-upload gap: 3 days; night share among European users: 15.6%).

Privacy. Demographic statistics describe adult patients only — no sign-up is allowed under the age of 18. All statistics are aggregates over the full cohort; no individual-level data, images, report contents or quotes are published, and no statistic is derived from fewer than 50 individuals unless its sample size is explicitly stated (English report corpus: n = 89 text documents). Ages are reported in 10-year bands. Geography is reported at no finer than country level. Scans and reports were processed on ReadYourLab’s existing analysis infrastructure under the same terms users already accepted; nothing was shared with third parties for this research.

Known limitations. Local time is approximated with a single time zone (US Central) for the North American cohort. DICOM header fields (scan date, patient birth year, sex, study description) are entered by imaging facilities and inherit their errors; 0.4% of studies had unparseable scan dates and were excluded from timing statistics. Anatomy classification is keyword-based and leaves 11% of studies unclassified. Flesch-Kincaid readability is defined for English; non-English reports were profiled for language and structure only.

Citing this report. Data, charts and statistics on this page are free to republish with attribution to “ReadYourLab Research” and a link to this page. For the underlying methodology, chart data or comment, contact peter@readyourlab.com.

About ReadYourLab

ReadYourLab is an AI-powered service that explains CT, MRI and X-ray imaging to the patients it belongs to, in plain language, in seven languages. It was co-founded by Peter Nemeth and Dr. Zoltan Nyarady MD, DMD, PhD. ReadYourLab is an educational tool, not a medical device, and does not provide diagnosis; every analysis carries that statement, and this research carries it too.