How patients write portal messages may decide whether their physician ever sees them

Patients who opened a portal message with their physician's name were far more likely to hear back from that physician than patients who opened with no greeting at all, according to an analysis of 3,619,390 patient messages published Aug. 24 in JAMA Network Open.

Messages that led with the clinician's last name drew a response from that clinician 38.7% of the time. Messages with no greeting drew one 25.7% of the time.

The cross-sectional study covered messages sent by 511,020 adults to primary care physicians, nurse practitioners and physician assistants across Mass General Brigham from Jan. 1, 2021, to Dec. 31, 2023.

Researchers used natural language processing to code both what each message asked for and how it was written, then measured which of the two explained the gap in whether patients from historically marginalized groups heard back from the clinician they wrote to. Writing style explained roughly half of it, and message content explained almost none.

Message writing style "may be a potential source of bias in message triage," said lead author Mitchell Tang, Ph.D., assistant professor of health policy and management at Columbia University Mailman School of Public Health.

Where the messages actually go

In most of the practices studied, a portal message does not arrive in the physician's inbox. It goes to a shared pool, where a triaging nurse or medical assistant reads it, weighs urgency and complexity and decides whether to answer it or forward it on.

The researchers measured two outcomes against that workflow: whether a thread drew any care team response, and whether it drew one from the clinician the patient originally wrote to. Across the full sample, 77.5% drew the first and 32.0% the second.

After adjusting for patient clinical characteristics, the recipient practice and clinician, and the time the message was sent, Black patients were 1.1 percentage points less likely than White patients to receive any care team response, a 1.4% relative decrease. On response from the intended clinician the difference was 3.7 percentage points, an 11.6% relative decrease.

Similar patterns held for patients with a preferred language other than English, patients whose highest education level was high school and patients covered by Medicaid or Medicare-Medicaid dual coverage.

Content did not explain the gap

The researchers ran two sets of features against the same regressions on a random subsample of roughly 990,000 English-language threads. One coded content, sorting messages into 20 topics with a zero-shot classifier. The other coded style: word count, reading level, sentiment, expressive punctuation, polite language and message opening.

Content was significantly associated with response rates overall but explained little of the demographic differences, accounting for 5.7% of the intended-clinician gap for Black patients, 12.7% for patients with only a high school education and 10.1% for Medicaid patients. For Hispanic patients the contribution was negative, meaning content differences narrowed the gap rather than widened it.

Writing style accounted for 48.0% of the gap for Black patients compared with White patients, 34.9% for Hispanic patients, 60.5% for patients with a high school education compared with those holding a college degree and 42.8% for Medicaid patients compared with commercially insured patients.

Which features mattered

The opening line carried the most weight. About 35% of messages led with the target clinician's last name, and roughly 29% carried no greeting at all.

Tang said in a statement that starting a message with a line as ordinary as "Dear Dr. Tang" made it far likelier that Tang himself would answer.

The same held for length. Threads of 200 or more words reached the intended clinician 49.7% of the time, compared with 21.0% for messages of 25 words or fewer. The adjusted model put an 18.7-percentage-point penalty on messages of three words or fewer.

Question marks, exclamation points and positive sentiment all tracked with higher response rates. Reading level did not move in a straight line. Response peaked among messages written at a sixth- to ninth-grade level and fell off above and below it.

One result runs the other way. Messages containing the word "please" reached the intended clinician less often than messages without it, 26.8% compared with 33.8%. Those figures are unadjusted, and the pattern may track the kind of request that tends to open with "please" rather than politeness itself.

The features driving the gaps varied by group. Length and expressive punctuation contributed across every demographic dimension, with patients from marginalized groups more likely to send short messages without question or exclamation marks. Among patients grouped by education and insurance, the opening line and sentiment carried additional weight.

What it means for triage

The content findings track how triage is designed to work. Administrative requests drew the lowest rates of physician response, at 19.6% for refills. Clinical questions drew the highest, including 51.6% for nutrition, diet or weight, 50.2% for blood pressure and 44.4% for non-refill medication questions.

The authors argue that stylistic bias is more tractable than bias rooted in race or income, because writing style can be observed and changed. Triage staff are unlikely to recognize the pattern in their own decisions, so awareness alone may help.

Portal interfaces could prompt patients toward more structured messages. Artificial intelligence could summarize the clinical core of a message and filter out stylistic noise, though the authors note that AI trained on the same data can carry the same bias forward. Most AI drafting systems in primary care inboxes still route through a human decision.

The study has limits. It draws on a single academic health system, excludes the 1% of messages written in languages other than English, and cannot capture responses delivered by phone or in person outside the record.

It is also cross-sectional, so it cannot establish that writing style caused the difference rather than standing in for something the researchers did not measure. The authors say the consistency of the pattern across message topics and patient subgroups makes that alternative less likely.