Measurement-based care means routinely tracking symptoms with validated scales and using the scores to guide treatment. Trials show it improves outcomes, yet it remains underused because collecting and reviewing scores is tedious. AI’s most useful role in psychiatry may be making that tracking automatic, not diagnosing anyone.
If you were designing an AI product for psychiatry to win headlines, you would build something that diagnoses depression from a selfie, or predicts psychosis from social media posts. If you were designing one to improve outcomes next year, you would build something far more boring: a system that makes sure every patient fills in a nine-question symptom scale before every visit, that the clinician actually looks at the trend, and that someone notices when the numbers stop improving.
That second product is measurement-based care, and it is the thesis of this piece. The quiet win for AI in mental health is not replacing clinical judgment. It is removing the friction that has kept one of psychiatry’s best-evidenced practices out of routine care for two decades.
If you are struggling and need help now, please contact your local emergency services or a crisis line. In the US, you can call or text 988.
What is measurement-based care in psychiatry?
Measurement-based care, or MBC, is the systematic use of standardized rating scales, such as the PHQ-9 for depression or the GAD-7 for anxiety, collected at regular intervals and reviewed during treatment to guide decisions. It is to depression what a blood pressure reading is to hypertension: not a diagnosis, but a number that tells you whether the treatment is working.
The idea is not new, and neither is the evidence. In a 2017 review in Psychiatric Services titled “A Tipping Point for Measurement-Based Care”, John Fortney, Jürgen Unützer and colleagues examined 51 studies and concluded that virtually every randomized trial providing frequent, timely feedback of patient-reported symptoms to clinicians during medication management or psychotherapy significantly improved outcomes.
Just as important was what did not work. The review found that one-time screening, infrequent measurement, and feeding scores back to clinicians outside the clinical encounter were ineffective. In other words, the number only helps if it arrives regularly and lands in front of the clinician at the moment of decision.
How much does measurement-based care improve outcomes?
The most striking single trial comes from Beijing. In 2015, Guo and colleagues at Beijing Anding Hospital published a randomized trial in the American Journal of Psychiatry comparing measurement-based care with standard care for 120 outpatients with moderate to severe major depression, using the same two antidepressants in both groups and outcome raters blind to the group.
The differences were large. In the measurement-based care group, 86.9% of patients responded to treatment, compared with 62.7% in standard care. Remission rates were 73.8% versus 28.8%. Time to remission was 10.2 weeks versus 19.2 weeks. The mechanism was visible in the data: clinicians in the MBC group made more treatment adjustments and reached effective doses sooner. Rates of dropout and side effects did not differ.
One trial in one setting should not be over-read, and the effect sizes in pragmatic settings are usually smaller. But the logic is compelling. Depression treatment often fails not because the right drug or therapy does not exist, but because clinicians wait too long to change a plan that is not working. A number that shows no improvement at week six is a prompt to act that a clinician’s impression of a pleasant conversation may not provide.
Accreditors agree. Since January 2018, the Joint Commission has required accredited behavioral health organizations in the US to monitor outcomes using a standardized instrument under its standard CTS.03.01.09.
Why don’t more clinicians use measurement-based care?
Because it is administratively annoying, and in mental health, administrative annoyance is fatal.
Someone has to send the questionnaire, chase the patient who did not complete it, enter the scores, graph them, and put the graph where the clinician will see it. Paper forms get lost. Portal questionnaires go unanswered. Scores land in a separate tab of the electronic record that nobody opens during a 20-minute visit. Clinicians who already feel buried in documentation, a problem we examined in The EHR Is Not the Burden. It Is Where Everyone Else’s Demands Land, experience MBC as one more form.
There is also a cultural objection. Some therapists feel that reducing a person’s suffering to a score undermines the therapeutic relationship. That concern is understandable, though the evidence suggests patients generally find routine measurement acceptable, and many find it validating to see progress on paper.
The Joint Commission requirement helped, but compliance can be superficial: organizations collect scores to satisfy auditors without feeding them into decisions. That is exactly the pattern the Fortney review found ineffective.
Where AI actually helps measurement-based care
This is where modern AI fits, and it fits best in unglamorous places.
Collection. Conversational tools can administer validated scales in a friendlier way than a form, by text message or chat, at a time that suits the patient, with reminders. The UK offers a relevant example at the front door of care. In a 2024 study in Nature Medicine, researchers led by Habicht, working with Limbic, examined Limbic Access, an AI chatbot used for self-referral to NHS Talking Therapies services. Across 129,400 people visiting 28 services’ websites, services using the chatbot saw a 15% increase in referrals compared with 6% for services using standard web forms, with larger increases among ethnic minority groups and non-binary people. The chatbot’s job was not to diagnose. It was to make the intake questions easier to complete.
Summarization. Language models can turn a series of scores and free-text comments into a one-paragraph summary at the top of the visit note: PHQ-9 down from 19 to 14 over eight weeks, sleep improved, anhedonia unchanged, patient mentioned stopping medication for three days. That puts the number where the Fortney review said it must be, in front of the clinician during the encounter.
Flagging. Simple rules, not sophisticated AI, can flag patients whose scores have not improved by a set threshold after a set time, prompting a treatment review. This is where the Guo trial found its benefit: faster adjustment.
Aggregation. Across a clinic, scores show which clinicians, programs or patient groups are getting results, which is the quality improvement layer the Fortney review saw as a secondary gain.
Notice what is missing from that list: AI deciding what the diagnosis is or which drug to prescribe. The clinical judgment stays human. The AI does the clerical work that has kept the practice from spreading.
A worked example: one clinic’s depression caseload
Picture a hypothetical community mental health clinic with 600 patients in active treatment for depression.
Before automation, staff try to collect PHQ-9 scores on paper at check-in. About a third of visits have a score. Clinicians rarely look at trends because they would have to find old forms.
After adopting an automated system, patients receive a text two days before each visit with the PHQ-9 and GAD-7. A short summary with the trend appears at the top of the note. Any patient whose PHQ-9 has not fallen by at least a quarter after eight weeks is flagged for review, and any positive answer on the self-harm item triggers same-day follow-up by a clinician under an agreed protocol.
Suppose completion rises to around 80%. The clinic can now see that roughly 150 patients are not improving after two months. Some need a dose change, some a different therapy, some a psychiatric review, some simply a conversation about whether they are taking their medication. None of those decisions was made by AI. All of them depended on AI making the data appear without extra work.
The self-harm item deserves care: automating it creates an obligation to respond reliably, including outside office hours or with clear messaging to patients about response times and crisis resources. A system that collects risk disclosures nobody reads is worse than one that does not ask.
The strongest counterargument
Critics raise a fair objection: scales can become the goal. A clinician under pressure to show improving scores may, consciously or not, treat the number rather than the person. Patients may learn to answer strategically, whether to seem better and be discharged or to seem worse and keep services. And scales miss things that matter enormously, such as meaning, relationships and functioning.
There is also a risk that automation makes measurement feel like surveillance, especially if scores are shared with payers to justify or limit sessions.
These objections argue for how MBC is used, not against it. Scores should open a conversation, not close one. They should be shared with patients, who can then discuss them. They should never be the sole basis for payer decisions. And they should be supplemented with measures of functioning and patient goals. Used this way, measurement adds a voice to the room rather than replacing one.
What this means for health systems, clinicians and vendors
For health systems, MBC is one of the few mental health AI investments with trial evidence for the underlying practice. The procurement question is not “how clever is the AI?” but “does it raise completion rates and put the trend in front of the clinician at the visit?” Measure both.
For clinicians, the practical ask is small: look at the trend, share it with the patient, and act when it stalls. As Mental Health in Primary Care Works. We Keep Paying for It the Wrong Way. argued, collaborative care models that rely on measurement already work in primary care. They are underused because of how they are paid for, not because of evidence.
For vendors, the lesson from remote monitoring applies, as we noted in Remote Patient Monitoring Pays for Data. Patients Benefit From Responses: selling data collection is easy, but outcomes come from what happens when the data arrive.
Frequently asked questions
What is an example of measurement-based care?
A common example is a patient with depression completing the nine-question PHQ-9 before each appointment. The clinician reviews the trend with the patient, and if scores have not improved meaningfully after several weeks, they adjust the treatment, for example by changing the dose, switching medication or adding therapy. The key is regular measurement used to guide decisions.
Does measurement-based care improve outcomes?
Yes, according to multiple trials. A 2017 review in Psychiatric Services found virtually all randomized trials providing frequent, timely symptom feedback to clinicians improved outcomes. A 2015 American Journal of Psychiatry trial found remission rates of 73.8% with measurement-based care versus 28.8% with standard care for major depression.
How can AI support measurement-based care?
AI can make symptom questionnaires easier to complete through chat or text, send reminders, summarize trends in the clinical note, and flag patients who are not improving. A 2024 Nature Medicine study found an AI self-referral chatbot increased referrals to NHS Talking Therapies services. The clinical decisions remain with clinicians.
Is measurement-based care required?
In the US, the Joint Commission has required accredited behavioral health organizations to monitor outcomes using a standardized instrument since January 2018, under standard CTS.03.01.09. Requirements differ elsewhere. Accreditation rules require measurement, but the evidence shows benefit only when scores are reviewed regularly and used in treatment decisions.
Can AI diagnose depression?
AI tools are being researched for detecting signs of depression from speech, text or behavior, but none should replace clinical assessment. Validated questionnaires like the PHQ-9 screen for symptoms and track severity. A diagnosis requires evaluation by a qualified clinician who considers history, context and other possible causes. Anyone concerned about depression should consult a clinician.
The takeaway
Psychiatry has had a proven practice sitting on the shelf for twenty years because it was too tedious to do routinely. If AI’s first real contribution to mental health is doing the tedium, that will be worth more than any algorithm that claims to read minds.
Does your service use symptom scores in every visit, and what finally made it stick, or what stopped it?
