Industries / Health / Healthcare / Behavioral health
Screening and assessment
Finding out who may have a mental disorder and naming it, with questionnaires, interviews and the diagnostic manuals that define each disorder.
7
Milestones achieved, from 1917 to 2026.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1917
A questionnaire for soldiers
Objective personality testing, in which people answer fixed written questions, began with Woodworth's Personal Data Sheet in 1917. It was built to pick out soldiers likely to break down under enemy bombardment in the First World War. Many similar tests soon followed in industry, where they were used to screen out job applicants judged maladjusted.
LessonA screening test spreads to uses it was never tested for.
Credit
Robert S. Woodworth
1973
Fake patients fool hospitals
David Rosenhan reported in Science that eight healthy people told US psychiatric hospitals they heard voices, were all admitted, mostly with suspected schizophrenia, and stayed 7 to 52 days although they reported no symptoms after admission. The paper helped persuade the American Psychiatric Association to rewrite its diagnostic manual. In 2023 the historian Andrew Scull, drawing on records found by the journalist Susannah Cahalan, called the study "a spectacularly successful case of scientific fraud".
LessonA famous study can steer a field for decades before anyone checks its records.
Credit
David Rosenhan; Susannah Cahalan and Andrew Scull, who checked it
1980
Diagnosis by written criteria
The third edition of the American Psychiatric Association's Diagnostic and Statistical Manual of Mental Disorders (DSM-III), published in 1980, gave each disorder operational criteria, a list of symptoms and rules for how many must be present, and checked how often clinicians agreed. Robert Spitzer chaired the task force from 1974. When the same method was tested for DSM-5 in 2013, two clinicians interviewing the same patient at 11 US and Canadian centers agreed well on most diagnoses, but 6 diagnoses reached only questionable agreement and 3 fell in the unacceptable range.
LessonWritten criteria raise agreement between clinicians but do not guarantee it.
Credit
Robert L. Spitzer and the APA task force; D. A. Regier and colleagues for the DSM-5 field trials
2001
Nine questions for depression
The PHQ-9 asks patients to rate nine symptoms from "not at all" to "nearly every day". In 6,000 patients at 8 primary care and 7 obstetrics and gynecology clinics, a score of 10 or more found major depression with 88% sensitivity and 88% specificity against an interview by a mental health professional. The questionnaire and its translations may be reproduced and distributed without permission. In 2019 an independent pooling of 58 studies and 17,357 people confirmed that a cutoff of 10 works best, with 88% sensitivity and 85% specificity against clinician-led interviews.
Credit
K. Kroenke, R. L. Spitzer and J. B. Williams; B. Levis, A. Benedetti, B. D. Thombs and the DEPRESSD Collaboration
2016
Neighbors split on screening
In 2013 the Canadian Task Force on Preventive Health Care recommended not routinely screening adults for depression in primary care, reversing its 2005 advice, because no trial had shown that screening helps. In 2016 the US Preventive Services Task Force recommended screening all adults, including pregnant and postpartum women, but only "with adequate systems in place to ensure accurate diagnosis, effective treatment, and appropriate follow-up".
LessonA screening test is only worth giving where diagnosis and treatment follow.
Credit
Canadian Task Force on Preventive Health Care; US Preventive Services Task Force
2024
A world manual for clinicians
The World Health Assembly approved the 11th revision of the International Classification of Diseases (ICD-11) in 2019, and it came into effect in 2022 as a fully digital system of about 17,000 codes. In March 2024 the World Health Organization (WHO) published its diagnostic manual for mental disorders under ICD-11, developed and tested with thousands of clinicians worldwide. In one of its field trials, 873 clinicians from around the globe diagnosed written cases slightly more accurately with the ICD-11 guidelines than with ICD-10. WHO's ICD-11 page counts 35 countries using ICD-11.
Credit
World Health Organization; G. M. Reed and the field trial teams
2026
A voice screener goes open
Kintsugi, a Berkeley company, built software that listens to how a person speaks, not what they say, for signs of depression and anxiety. In its own staff's study of 14,898 adults, 25 seconds of speech flagged a PHQ-9 score of 10 or more with 71.3% sensitivity and 73.5% specificity. It worked with the US Food and Drug Administration (FDA) from 2022 toward De Novo authorization, a route for new kinds of device. On 10 February 2026 it said it was shutting down after spending $30 million, and released its Depression-Anxiety Model, built on OpenAI's Whisper speech model, on Hugging Face under the Apache 2.0 license with scores, labels and demographics for its validation and test recordings.
LessonA screening tool can run out of money before it reaches regulatory clearance.
Credit
Kintsugi (Grace Chang, founder)
Watch for
A redesigned DSM
A committee of the American Psychiatric Association proposes renaming the DSM the Diagnostic and Scientific Manual and organizing it in four parts: contextual factors such as income and culture, biomarkers and biological factors, diagnoses at different levels of specificity and severity, and features shared across many conditions. The proposal appeared as five papers in the American Journal of Psychiatry, with no timeline for a new edition.
Achieved whenthe APA publishes an edition built this way, with field trials showing how often two clinicians agree on its diagnoses.
Claimed by
the APA's Future DSM Strategic Committee, chaired by Maria Oquendo, January 2026
Open
01
Check an open voice screener
An outside report card for Kintsugi's open Depression-Anxiety Model. Part one uses only the published scores: sensitivity and specificity at a PHQ-9 cutoff of 10 by age band, gender and race or ethnicity, with thresholds set on the validation split and tested on the test split. Part two feeds the model synthetic speech only: the same neutral scripted passages read by many text-to-speech voices, and the same audio passed through phone codecs and background noise, to measure how much the score moves when only the voice or the line changes. No real person's voice is assessed.
The briefCredit · unclaimed02
Test the questionnaire checkers
A public audit of how free online PHQ-9 and GAD-7 checkers, and the versions built into apps, score answers. The builder writes a fixed set of fictional answer patterns, such as totals just below and above each published cutoff (5, 10, 15 and 20 for the PHQ-9), a skipped question, and a positive answer to question 9, which asks about thoughts of self-harm. Each pattern is entered into each tool, and the total, the severity label and what the tool shows after a positive question 9 are recorded and compared with the published scoring instructions.
The briefCredit · unclaimed
Open 01 · the brief
Check an open voice screener
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An outside report card for Kintsugi's open Depression-Anxiety Model. Part one uses only the published scores: sensitivity and specificity at a PHQ-9 cutoff of 10 by age band, gender and race or ethnicity, with thresholds set on the validation split and tested on the test split. Part two feeds the model synthetic speech only: the same neutral scripted passages read by many text-to-speech voices, and the same audio passed through phone codecs and background noise, to measure how much the score moves when only the voice or the line changes. No real person's voice is assessed.
Why it is still open
The model and about 15,000 scored recordings with labels and demographics have been public since February 2026, and the model card warns that biases related to language, accent or demographics may be present. The only published accuracy study is by Kintsugi's own staff. The model's Hugging Face page shows three user questions about installation and output and no outside evaluation, and a web search in November 2026 found none.
Built with what exists
the KintsugiHealth/dam model and dam-dataset on Hugging Face (both Apache 2.0) and the threshold-tuning code that ships with the model; Python; an open text-to-speech model such as Kokoro-82M (Apache 2.0) or commercial voices; ffmpeg to pass audio through telephone codecs.
First step
Load the two splits, reproduce the model card's threshold example with 20% of recordings left undecided, and break the result down by gender.
Done looks like
A table of sensitivity and specificity, each with a 95% confidence interval, for every subgroup with at least 200 recordings in the test split, beside the overall figure. A synthetic-voice test of at least 50 voices reading the same 20 passages, reporting the share of passages whose severity level changes with the voice alone. Code and synthetic audio published, and a second person reruns the analysis and gets the same table.
Open 02 · the brief
Test the questionnaire checkers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public audit of how free online PHQ-9 and GAD-7 checkers, and the versions built into apps, score answers. The builder writes a fixed set of fictional answer patterns, such as totals just below and above each published cutoff (5, 10, 15 and 20 for the PHQ-9), a skipped question, and a positive answer to question 9, which asks about thoughts of self-harm. Each pattern is entered into each tool, and the total, the severity label and what the tool shows after a positive question 9 are recorded and compared with the published scoring instructions.
Why it is still open
Anyone may reproduce, translate, display or distribute the PHQ and GAD-7 without permission, so copies are everywhere. Searches of PubMed and the web in November 2026 found no published audit of how these copies score answers.
Built with what exists
the PHQ and GAD-7 forms and scoring instructions from the PHQ Screeners site; a browser automation tool such as Playwright, or a browser agent such as Browser Use (see the Note "Deleting 80% to start over"); a spreadsheet.
First step
Write the 12 answer patterns with the expected score and label for each, and run them through five of the most visited checkers.
Done looks like
At least 50 tools each tested with 12 patterns. The share of tools that miscount totals, use cutoffs other than the published ones, change the score of a skipped question without saying so, or show no help information after a positive question 9. Two people enter every pattern independently and agree on at least 95% of records. The audit is repeated after six months.