Industries / Health / Healthcare / Medical labs
Reporting results
Getting each result, with its units and normal range, to the person who has to act on it, and making sure the dangerous ones are noticed in time.
6
Milestones achieved, from 1972 to 2014.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1972
Dangerous results phoned at once
In 1969 a young man in a coma at the Los Angeles County–University of Southern California Medical Center had a blood sugar of 6 mg/dL, a life-threatening low. The printed result reached the ward within two hours, but nobody read it until the next morning, and he suffered permanent brain damage. The lab's response was to phone a short list of "panic values" (later "critical values") straight to a doctor. George Lundberg published the list in 1972; by his account labs across the US adopted their own versions within weeks, and the College of American Pathologists made such a system a condition of accreditation.
LessonA result has not been reported until someone who can act on it knows it.
Credit
George D. Lundberg, LAC–USC Medical Center
1994
One code for each test
Clem McDonald of the Regenstrief Institute in Indianapolis started LOINC (Logical Observation Identifiers Names and Codes), a catalog that gives each lab test one universal code, so that results sent from a lab's computer to a hospital or insurer mean the same thing at both ends. By LOINC's own count it has registered users in nearly every country and is a named national standard in more than 30.
Credit
Clem McDonald and the LOINC Committee, Regenstrief Institute
2009
Abnormal results not passed on
Researchers reviewed the records of 5,434 patients aged 50 to 69 in 23 US primary care practices. For 7.1% of abnormal results (135 of 1,889), there was no sign that the patient had been told; the rate ran from 0% to 26.2% between practices. Practices with better routines for handling results had fewer failures. Practices whose records were half on paper and half electronic did worse than practices with no electronic records at all.
LessonA partly computerized process can lose more results than a paper one.
Credit
Lawrence P. Casalino and colleagues (Archives of Internal Medicine, 2009)
2012
Results by text message
In Zambia's Southern Province, dried blood spots from infants are sent to regional labs to be tested for HIV, and results used to come back by post. Project Mwana's software sent each result from the lab to the clinic as an SMS text instead. At 10 clinics, the mean time for a result to reach the clinic fell from 44.2 to 26.7 days, and the time to reach the child's caregiver fell from 66.8 to 35.0 days. Only 0.5% of the texted reports checked differed from the paper copies.
Credit
Patrick Seidenberg, Donald M. Thea and colleagues, Boston University Center for Global Health and Development (Bulletin of the World Health Organization, 2012)
2014
Patients get their lab reports
A US rule from the Department of Health and Human Services changed the lab law, the Clinical Laboratory Improvement Amendments (CLIA), and the health privacy rule, HIPAA, so that patients can get their completed test reports directly from the lab on request. It overrode state laws that barred labs from releasing reports directly to patients or without the ordering doctor's consent. The rule took effect on 7 April 2014, and labs covered by HIPAA had to comply by 6 October 2014.
Credit
US Department of Health and Human Services (Centers for Medicare & Medicaid Services, Centers for Disease Control and Prevention, Office for Civil Rights)
2014
Software releases most results
Autoverification is software that checks each result against rules set by the lab director, such as limits, changes from the patient's last result and signs of a damaged sample, and releases it with no person reviewing it. The core chemistry lab at the University of Iowa Hospitals and Clinics, a 734-bed hospital, raised the share of results released this way from 40% to 99.5% over 13 years, including a high share of critical values.
Credit
Matthew D. Krasowski and colleagues, University of Iowa (Journal of Pathology Informatics, 2014)
Watch for
Labs explain results with AI
Quest Diagnostics' AI Companion, built on Google's Gemini models, reads up to five years of a person's Quest results in its MyQuest app and explains the values and trends. Labcorp's MyLabcorp app does the same job with OpenAI's models. Neither company's announcement gives accuracy figures. Labcorp's own survey found 41% of consumers using AI to interpret lab results.
Achieved whenan outside group scores the explanations on a fixed, published set of reports, including how often they tell the user to contact a clinician the same day about a critical value.
Claimed by
Quest Diagnostics, March 2026; Labcorp, May 2026
AI drafts result notes
Stanford Health Care built a tool on Anthropic's Claude 3.5 Sonnet that drafts the message a doctor sends a patient to explain lab, imaging and pathology results. In a pilot from January to March 2025, 244 primary care clinicians used it. Of the 93 who answered a survey, 72.0% said it gave higher-quality explanations, and the median time they thought it saved was 1.1 minutes per result. Their main complaints were accuracy and completeness, and the study measured opinions, not errors.
Achieved whena study outside Stanford counts errors in the drafts and measures clinicians' time against a group without the tool.
Claimed by
Stanford Health Care and Stanford University School of Medicine, August 2025
Open
01
An atlas of normal ranges
A public, machine-readable table of the reference ranges (the "normal" band printed beside each result) that at least 200 labs in at least 15 countries publish for 30 common tests. Each row records the units, converted to one system, the sample type (serum or plasma), any age and sex bands, the page it came from and the date it was read. Then run a fixed set of made-up results through every lab's ranges and count how often the same number would be flagged high, low or normal.
The briefCredit · unclaimed02
Staged tests of AI callers
An AI voice agent that phones critical results from made-up lab reports to actors playing ward nurses. It must get a read-back (the listener repeats the patient's name and the value), escalate when no one answers, and log each step. Score it on scripted surprises, such as a wrong patient, a misheard number or a busy line, and run the same scripts with human lab staff for comparison.
The briefCredit · unclaimed
Open 01 · the brief
An atlas of normal ranges
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public, machine-readable table of the reference ranges (the "normal" band printed beside each result) that at least 200 labs in at least 15 countries publish for 30 common tests. Each row records the units, converted to one system, the sample type (serum or plasma), any age and sex bands, the page it came from and the date it was read. Then run a fixed set of made-up results through every lab's ranges and count how often the same number would be flagged high, low or normal.
Why it is still open
Studies so far cover one country or one test at a time. In 2026, Glen Hortin and colleagues searched 60 labs' online test directories, mostly in the US, and found 14 different normal ranges for serum potassium; only 4 labs gave serum and plasma separate ranges, although the two usually differ by 0.2 to 0.4 mmol/L. AI tools that explain results lean on these ranges. I found no public table that spans many tests and countries.
Built with what exists
an AI agent that browses web pages and pulls out tables, such as Browser Use (see the Note "Deleting 80% to start over") or Claude or Gemini with web tools; LOINC codes to name each test; NLM's open ucum-lhc library to convert units; Python for the flag counts.
First step
Collect the potassium, sodium, ALT, TSH and hemoglobin ranges from 20 labs' public test directories in three countries, and count the flips for 100 made-up values.
Done looks like
At least 200 labs in at least 15 countries and 30 tests, every row linked to its source page. A second person rechecks a random 10% of rows and agrees on at least 98%. A published count, per test, of how many of 1,000 made-up values change flag between labs. Data and code under an open license with a DOI.
Open 02 · the brief
Staged tests of AI callers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An AI voice agent that phones critical results from made-up lab reports to actors playing ward nurses. It must get a read-back (the listener repeats the patient's name and the value), escalate when no one answers, and log each step. Score it on scripted surprises, such as a wrong patient, a misheard number or a busy line, and run the same scripts with human lab staff for comparison.
Why it is still open
In the College of American Pathologists' 2008 study of 3,545 critical-result calls at 121 labs, staff took a median of 4 minutes from result to call, and 96% of calls included a read-back; one institution used an automated system that confirmed contact. A 2012 review for the US Centers for Disease Control and Prevention found the evidence for automated notification insufficient to recommend for or against it. Vendors now advertise AI callers for critical results with acknowledgment codes but cite no test. I found no published test of an AI caller.
Built with what exists
a speech recognition model (see the Note "Fewer word errors"); a language model with tool calls, such as Claude, Gemini or an open-weight model (see the Note "A trillion open parameters"); a telephone API such as Twilio; made-up results from Synthea, MITRE's open generator of synthetic patient records; a hospital lab's published critical-value list. The advance that makes this possible is voice models that hold a phone conversation in real time.
First step
Write 20 call scripts from a published critical-value list and run them with one actor and a prototype agent on a test phone line.
Done looks like
200 staged calls with at least 10 actors, each recorded with consent and scored by two people with lab training for the correct patient and value read back, time to a completed notification, and escalation when no one answers. The same 200 scripts run with human callers. Scripts, scores and recordings published.