Precision Neuropsychiatry

Established Precision neuropsychiatry is the claim that a measurement taken from an individual patient can tell you which treatment that patient should get. Four programmes make the claim in different currencies: closed-loop and individually targeted brain stimulation, psychedelic- and ketamine-assisted treatment, digital phenotyping from passive sensor data, and biomarker or pharmacogenomic stratification. Established They are usually reviewed separately, by different specialists, in different journals. This brief compares them against one standard: what does a blinded, adequately powered, prospectively registered trial show?

Established Judged that way, the four programmes are not at comparable stages, and the differences are larger than the field’s own rhetoric suggests. Individually targeted stimulation has the single strongest sham-controlled result in the set and the smallest sample behind it. Established Psychedelic-assisted treatment has the largest effect estimates in the set and the worst blinding in clinical research. Frontier Digital phenotyping has abundant within-cohort accuracy and a documented failure to generalise between cohorts. Frontier Biomarker stratification has the most funding, the longest history, and no prospectively validated treatment-assignment rule in routine use.

Frontier One measured finding cuts across all four and is the most important thing in this brief. If patients truly differ in which treatment works for them, the spread of outcomes in a drug arm should exceed the spread in a placebo arm; the excess variance is the signature of real individual differences in response. Established Variance-ratio meta-analyses in both schizophrenia and depression have repeatedly found that ratio close to one. Frontier That does not prove personal response differences are absent, but it bounds them: whatever precision psychiatry is chasing, the trials run so far do not contain much of it.

Established This is a synthesis, and it does not repeat what neighbouring briefs own. Bioelectric Medicine holds the device-level record for neuromodulation, Neural Interfaces holds the electrode-longevity question underneath it, and Cognitive Enhancement holds the effect sizes for intervening in healthy brains. Established What none of them owns is the joint clinical question — whether psychiatry can assign treatments by measurement at all — and that question is this brief’s subject. Nothing here is clinical guidance; it is a reading of the trial record.

Established A note on sourcing. This brief was commissioned in September 2026 from the Institute’s research base. Reading-list entries without links are cited from the bibliographic record rather than re-fetched, and claims are dated no later than early 2026 unless carried by a linked source.

Established Start with the baseline the precision claim has to beat. The largest network meta-analysis of antidepressants pooled 522 double-blind trials and 116,477 participants and found all twenty-one drugs more effective than placebo, with odds ratios spanning roughly 1.4 to 2.1 and standardised effects that are small by any other field’s standard. Established The pragmatic STAR*D sequence reported a cumulative remission figure that a 2023 patient-level reanalysis put at roughly half the originally publicised number, after correcting for protocol deviations in outcome measurement. Frontier The honest summary is that average treatments work modestly and that the field’s headline numbers have been revised downward when the underlying data were reopened.

Established Biomarker stratification has the longest run and the clearest set of negative results. The candidate-gene era ended with a test of eighteen historical depression genes in 621,214 individuals that found no support for any of them, including the interaction models that had dominated two decades of publication. Established Pharmacogenomic-guided prescribing has since been tested in large randomised trials: the best-known industry-sponsored trial missed its primary symptom endpoint while reporting positive secondaries, and a large independent trial in a veterans’ health system found remission effects that were small and did not persist. Frontier Primary endpoint missed, secondaries promoted, recurs often enough in this literature to be structural rather than accidental.

Frontier Imaging-derived subtypes are the most cited and least reproduced part of the programme. A 2017 study reported four resting-state connectivity biotypes of depression in more than a thousand participants and claimed they predicted response to stimulation. Established A 2019 reanalysis using a comparable pipeline failed to recover the clusters, and reported that the apparent subtypes were consistent with a continuous distribution. Established More generally, a 2024 analysis of clinical prediction models built on schizophrenia trial data found that models achieving useful discrimination within the trial that trained them fell to chance when applied to a different trial. That is the central engineering fact about psychiatric prediction: within-sample accuracy is cheap and does not transfer.

Established Stimulation is where precision has actually been demonstrated, and the demonstration is in neurology rather than psychiatry. Adaptive deep brain stimulation for Parkinson’s disease reads a physiological signal from the same electrode that stimulates and adjusts output in response; a blinded randomised feasibility trial reported a substantial reduction in time spent with the patient’s most bothersome symptom against conventional stimulation, and a regulator cleared a commercial adaptive system in February 2025. Established The reason it works is legibility: the beta oscillation is a stable, measurable correlate of the motor state. Frontier Psychiatry has no equivalent. The closest published attempt is a single-patient closed-loop implant for treatment-resistant depression in which a personalised amygdala biomarker triggered stimulation, reported in 2021 with an n of one.

Established Individually targeted accelerated stimulation gives the strongest sham-controlled psychiatric effect in the set, from a very small trial. A double-blind randomised trial of an accelerated, functional-connectivity-targeted theta-burst protocol in twenty-nine participants reported remission in roughly 79 per cent of the active arm against roughly 13 per cent of sham at the primary timepoint. Frontier That is a large effect from a single site with fewer than thirty participants, and the literature’s open questions are durability beyond the first weeks and whether the individualised targeting, the acceleration, or the total dose delivered is doing the work. Frontier Multi-site replication is the item the field needs and does not yet have at scale.

Established Implanted stimulation for depression has an unusual evidentiary shape: negative blinded trials, positive open-label follow-up. Two industry-sponsored randomised trials at different targets were stopped or reported as failing to separate from sham at their primary timepoints, while long-term open-label cohorts from the same programmes report high response rates sustained over years. Frontier Both can be true — slow-onset effects are poorly served by short blinded windows — and the divergence is also exactly what unblinded outcome assessment produces. No trial has been designed to separate the two explanations.

Established Psychedelic-assisted treatment has the largest reported effects and a blinding problem that is not a technicality. In a 2022 randomised trial of a synthetic psilocybin formulation in treatment-resistant depression with 233 participants, the high-dose arm separated from the comparator at three weeks by several points on a standard depression scale, with the difference attenuating by twelve weeks. Established Adverse-event reporting in that trial included suicidal ideation and self-injurious behaviour, recorded as trial endpoints; sponsors and independent commentators have disagreed about their interpretation and base rates in this population. Established A 2021 head-to-head trial of psilocybin against a standard antidepressant did not separate on its primary endpoint.

Established Blinding failure in this literature has been measured, not merely alleged. Analyses of psychedelic trials report that participants and raters guess assignment correctly at rates far above chance, in some trials above ninety per cent. Established In the phase 3 programme for MDMA-assisted therapy, an advisory committee in June 2024 voted against both effectiveness and favourable benefit-risk, and the regulator issued a complete response letter in August 2024 requiring an additional trial. Frontier That regulatory outcome is the single most informative datum in the subfield, because it was reached on the evidence the sponsor considered its strongest.

Frontier Digital phenotyping has the weakest published effect sizes and the clearest commercial failure. Passive smartphone and wearable features correlate with symptom severity at magnitudes that are usually small, and the prediction models built on them have rarely been validated in an independent cohort. Established The best-funded company in the sector wound down its clinical operations in 2023, and a national regulator’s digital-health precertification pilot ended in 2022 without producing the streamlined pathway it was designed to test. Frontier The technology is not disproven; the measured claim that it identifies who will relapse, prospectively and out of sample, is not yet supported.

Frontier Is there any psychiatric biomarker that survives external validation? The requirement is specific: a measurement taken before treatment that predicts differential response between two treatments, in a cohort other than the one that discovered it, with the interaction pre-registered. Frontier Candidates exist — frontal theta asymmetry, rostral anterior cingulate activity, inflammatory markers, sleep and circadian variables — and none has cleared that bar in a trial designed for it.

Frontier Can blinding be repaired in trials of drugs with unmistakable acute effects? Proposals include active comparators chosen to mimic subjective effects, dose-masking across multiple arms, measuring expectancy at baseline and adjusting for it, and blinded independent raters who never meet the participant. Frontier Each has been tried somewhere; none has been shown to restore blind integrity in a psychedelic trial, and a field that cannot blind cannot cleanly estimate its own effect sizes.

Frontier Is the individual-variability premise true at all? Variance-ratio methods test it directly and have returned ratios near unity across several drug classes and disorders. Speculative The counter-argument is that these methods are underpowered for small subgroups and that cross-over designs would answer better; few cross-over trials in psychiatry are large enough to settle it. This is the most consequential open question in the field and the cheapest to attack with existing data.

Frontier Are symptom rating scales the wrong endpoint? Regulators accept clinician-rated scales, which have modest inter-rater reliability and compress heterogeneous experience into one number. Speculative Functional endpoints — employment, hospitalisation, social contact — are harder to fake and slower to move, and no regulator has accepted one as primary here.

Established Measurement is the first bottleneck and it sits under everything else. Psychiatric outcome measurement rests on clinician-administered scales with known reliability limits, applied to diagnostic categories defined by committee rather than by mechanism. Frontier A prediction model cannot be more reliable than its target label, and much of the apparent unpredictability of treatment response may be measurement noise rather than biology.

Established Statistical power for interaction effects is the second, and it is arithmetic rather than opinion. Detecting that a marker changes which treatment works requires roughly four times the sample needed to detect the main effect of the same size. Established Psychiatric trials are typically sized for main effects. A field that wants stratification and funds trials at main-effect sizes has pre-committed to inconclusive results.

Established Blinding is the third, and in one subfield it is close to total. An intervention with unmistakable acute subjective effects unblinds participants, and often raters, regardless of protocol. Frontier Measured guess rates above ninety per cent make the placebo arm a different experience rather than a control, and expectancy in this population is known to move rating-scale scores.

Frontier Generalisation is the fourth. Models trained in one trial degrade to chance in another, through site effects, differing inclusion criteria and differing rater training. Established Published prediction work in this field has usually reported internal cross-validation and not external validation, and the gap between those two numbers is the whole result.

Frontier Trial-population selection is the fifth and is rarely counted as a bottleneck. Trials in severe psychiatric illness commonly exclude participants at the highest levels of acute risk, so the evidence base is generated in a population narrower than the one the treatment is meant for. Established This is a documented feature of the literature, and it limits what any effect estimate, precise or not, can be said to apply to.

Frontier The programme depends on a measurement with test-retest reliability better than the outcome it predicts. Functional MRI connectivity measures, the workhorse of psychiatric biomarker work, have reliability that varies sharply with scan length and preprocessing, and short scans give estimates too unstable to support individual-level decisions. Frontier Longer acquisitions improve this materially, which is a solvable dependency and an expensive one.

Established It depends on device platforms that can sense as well as stimulate. Closed-loop operation requires an implant that records a signal while delivering current without saturating its own amplifier, which is a hardware property and not a software one. Established Bioelectric Medicine records the state of those platforms and the Parkinson’s results they have produced; this brief takes them as given and asks what psychiatry could do with them.

Frontier It depends on electrodes that survive long enough to be worth implanting for a chronic psychiatric indication. Neural Interfaces records the twenty-year human yield data and the finding that connector failure, rather than tissue response, dominated the founding dataset. Frontier A psychiatric indication is a decades-long commitment in a younger population than most implant cohorts, so longevity matters more here than in the applications that generated the data.

Established It does not depend on a new drug class, and treating it as though it does is a category error. Every question here can be asked of compounds already licensed or in phase 3. The binding constraints are design, endpoint and blinding, none of which requires a discovery.

Frontier The decisive experiment is a prospective, adequately powered randomised trial in which patients are assigned to one of two treatments by a pre-specified biomarker, with the biomarker-by-treatment interaction as the registered primary endpoint. Nothing in this field has been done to that standard, and almost everything in it is argued as though something had. Established The design is ordinary, the statistics are textbook, and the obstacle is that such a trial needs roughly four times the participants of a conventional two-arm trial. Frontier It could be run with existing markers and existing drugs; no sponsor has funded one at the required size, which makes this an unscheduled experiment rather than a hard one.

Frontier The second experiment is a blinding-integrity trial that treats unblinding as the outcome rather than as a footnote. Randomise participants across several masked conditions, collect assignment guesses and confidence from participants and raters at every visit, measure baseline expectancy, and report the effect estimate with and without adjustment. Frontier This has never been done as a primary-purpose trial in the psychedelic literature, and until it is, the effect sizes in that literature have an unquantified upper bound of expectancy contribution.

Established The natural experiment already running is the variance-ratio analysis of the existing trial archive. Every completed placebo-controlled trial with reported outcome dispersion contributes to it, and no new patients are needed. Established Published applications in schizophrenia and depression have returned ratios close to one, which is evidence against large personal differences in response. Frontier Extending it systematically across drug classes, stimulation and psychotherapy would either license the precision programme or bound it, using data that already exists.

Frontier The fourth is an adequately powered multi-site replication of accelerated individualised stimulation with an active sham and long follow-up. The single-site result is strong enough to justify it and small enough that it cannot settle anything on its own. Frontier A replication that also randomised targeting method against a standard anatomical target would separate the individualisation claim from the acceleration claim, which is the specific ambiguity the current evidence leaves open.

Established Closed-loop hardware is the part of this subject where the engineering is genuinely solved enough to deploy. Commercial implantable pulse generators now sense local field potentials, classify a state and modulate output within the therapeutic loop, with a regulator-cleared system on the market since February 2025. Frontier The psychiatric version needs a control signal, not better hardware, which is why the hardware has arrived in neurology first.

Frontier Targeting is an imaging and modelling problem with a real cost curve. Individualised functional-connectivity targeting for transcranial stimulation requires per-patient imaging, a processing pipeline and a neuronavigation system, which is roughly an order of magnitude more resource per patient than standard scalp-landmark targeting. Frontier Whether that cost buys outcome is exactly the question the multi-site replication would answer.

Established Digital phenotyping is engineering-cheap and validation-expensive. Collecting accelerometry, keyboard dynamics and location entropy is trivial on consumer hardware; the expensive parts are ground truth, retention, missing data and privacy-compliant storage. Frontier The failure mode is not sensing but labelling: the target variable is a periodic questionnaire, and the model inherits its noise.

Speculative Delivery capacity, not device count, is the constraint on the stimulation results if they hold. An accelerated protocol occupies a machine and a clinical team for days per patient, which sets clinic throughput. Handwave Claims that a strong small-trial result implies population-scale availability skip the staffing arithmetic.

Established The device record belongs to a neighbouring brief and this one relies on it. Bioelectric Medicine holds the sham-controlled neuromodulation results and the adaptive-stimulation trial record, including the divergence between the two sciences that share the name. Established This brief adds only the psychiatric reading of it: the closed-loop success is in a disease with a legible oscillatory biomarker, and psychiatry does not have one.

Established The electrode question belongs to another. Neural Interfaces holds the chronic-yield data that determines whether an implanted psychiatric indication is a twenty-year proposition or a five-year one. Frontier Its finding that connector failure dominated the founding dataset matters here because a psychiatric implant would be placed in patients with decades of life ahead.

Established The healthy-brain effect sizes belong to a third, and they set a prior. Cognitive Enhancement records that honest meta-analyses of improving healthy cognition return effects between 0.1 and 0.3 standard deviations, and that the largest body of stimulation work in that field has a mechanism problem. Frontier A field that struggles to move healthy function should be read sceptically when it reports very large effects in disordered function from very small samples.

Established The general lesson comes from outside neuroscience. Precision Medicine records that the wider field reliably predicts who a drug will harm and reliably fails to predict who it will help. Established Psychiatry fits that pattern exactly: pharmacogenomic panels are most defensible for metabolism and tolerability, and least defensible for choosing which agent will work.

Established The research framework that was supposed to fix this has not produced a treatment-assignment rule. A national funder reoriented much of its portfolio around dimensional constructs rather than diagnostic categories from around 2010, on the argument that the categories were the obstacle. Established Its own former director has said publicly that the large sums spent during that period did not move population outcomes on suicide, hospitalisation or recovery. Frontier The framework remains scientifically defensible and has not yielded the clinical decision rule it was justified by.

Established Regulatory designations have been widely misread as approvals. Breakthrough-device and breakthrough-therapy designations shorten review interaction; they are not findings of effectiveness, and several designated psychiatric programmes have subsequently failed or been returned for more evidence. Established The August 2024 complete response letter on MDMA-assisted therapy is the clearest instance, following an advisory vote against effectiveness two months earlier.

Frontier Access has moved ahead of federal evidence in specific jurisdictions, which creates a natural experiment. Two US states established regulated supervised-use programmes for psilocybin outside the drug-approval system, with service centres operating from 2023 onward, and an Australian regulator permitted authorised prescribers to supply MDMA and psilocybin for specified indications from July 2023. Frontier These schemes generate observational safety and outcome data at a scale trials will not reach, and their registries are the most valuable unexploited dataset in the subfield.

Established Reimbursement determines which precise things get done. Accelerated individualised stimulation is a multi-day, staff-intensive procedure, and coverage decisions rather than trial results set how many clinics offer it. Frontier No payer has a code that reimburses biomarker-guided treatment assignment as such, which means that even a validated marker would enter practice through the slowest available door.

Frontier Consent in an unblindable trial is a live problem rather than a formality. If participants can tell which arm they are in, the informed-consent document is describing a different experiment from the one being run, and expectancy becomes part of the intervention. Established Measured guess rates in the psychedelic literature are high enough that this is a design fact, not a hypothetical.

Frontier The therapy-versus-drug confound raises a fairness question about what is being licensed. Trials of psychedelic-assisted treatment bundle a pharmacological agent with many hours of structured psychological support. Established No completed trial has factorially separated the two components, so the evidence supports the bundle and not either part, and regulators licensing one part face an evidentiary gap that critics and proponents both acknowledge.

Established Trial-population exclusions shape who the evidence can speak for. Where trials exclude participants at the highest acute risk, reported adverse-event rates — including suicidality endpoints — are generated in a lower-risk sample than the clinical population. Frontier That is an argument for registry follow-up rather than against the treatments, and the jurisdictions running supervised-access schemes are the ones positioned to supply it.

Frontier Passive sensing is surveillance with a clinical justification, and the justification is currently weak. Digital phenotyping collects location, movement and communication metadata continuously, which is among the most sensitive data any medical technology gathers. Frontier The published predictive gain over periodic questionnaires is small, so the privacy cost is being incurred ahead of a demonstrated benefit.

Established Commercial pharmacogenomic testing is marketed beyond its trial evidence. Panels are sold on the basis of trials whose primary endpoints were missed, and the promotional framing rarely distinguishes metabolism predictions, which are better supported, from efficacy predictions, which are not. This is where the gap between evidence and claim reaches patients fastest.

Established The scale of the problem is what makes small effects matter. Depressive and anxiety disorders are among the largest contributors to years lived with disability worldwide, and the treatments in routine use have small average effects with high non-response. Frontier A stratification rule that raised the first-treatment success rate even modestly would be worth more, in aggregate disability terms, than most of what the field currently pursues.

Frontier If the variance-ratio result holds, the field’s strategy is misallocated rather than merely slow. Little excess variance in drug arms implies that the largest available gains come from better average treatments, faster access and better measurement, not from sorting patients. Speculative That is an uncomfortable conclusion for a research economy organised around individual difference, and it is testable now.

Speculative The reputational stake is larger than any single treatment. Psychiatry has absorbed two public reversals in a generation — the candidate-gene literature and the simplest chemical-imbalance account — and a third would affect recruitment, funding and public trust in psychiatric evidence generally. Frontier Fields recover from reversals they announce themselves and badly from ones announced by others.

Handwave Claims that measurement will dissolve psychiatric diagnosis belong here, flagged. The strong version holds that biomarkers will replace the diagnostic manual with mechanism-defined disorders. Handwave Fifteen years of dimensional research has produced constructs, instruments and a large literature, and no biologically defined psychiatric disorder is in clinical use. The claim may still be right; it is not currently supported by a result.

These horizons track what would have to be demonstrated for a claim to change status, not predictions of arrival.

• 10 yr: Frontier A multi-site replication of accelerated individualised stimulation with an active sham and durable follow-up is affordable today and is the single most likely status change on this horizon. Frontier A biomarker-stratified trial with the interaction as its registered primary endpoint is equally affordable and currently unfunded; its absence ten years from now would say more about the field than any statement it makes. Frontier Registry outcomes from the supervised-access jurisdictions will accumulate whether or not anyone designs them well.
• 25 yr: Speculative If any marker clears external validation, this is the horizon on which reimbursement, rater training and clinical workflow decide whether it is used, and those are slower than the science. Speculative Closed-loop psychiatric stimulation becomes plausible only if a control signal is found; without one, implanted psychiatric devices remain open-loop and are judged by blinded trials they have so far not passed.
• 50 yr: Speculative A mechanism-defined psychiatric nosology, if it arrives, arrives here rather than sooner, and would be recognisable by a diagnosis that changes treatment assignment rather than merely renaming a syndrome. Handwave Predictions that this happens because data volume increases assume that the limiting factor is data rather than label reliability, which the external-validation results contradict.
• 100 / 250+ yr: Handwave Claims about the elimination of psychiatric illness through individualised intervention belong here. Handwave They require a causal model of the disorders that does not exist, endpoints nobody has validated, and an assumption that variability between patients is biological rather than measured, which is the one premise the existing data most directly challenges.

• Depends on This brief sits downstream of three on the map. It waits on the device record in Bioelectric Medicine, specifically the sensing-capable implant platforms that made adaptive stimulation possible in Parkinson’s disease; on the chronic-yield results in Neural Interfaces, which determine whether an implanted psychiatric indication is viable over decades; and on the effect-size priors in Cognitive Enhancement, which bound what stimulation plausibly does to a brain at all. None of those three supplies the thing this brief most needs, which is a validated psychiatric control signal.
• Requires (not on this map) First, a pre-treatment measurement that predicts differential response in a cohort other than the one that discovered it, which is the thing the whole programme is named after and does not have. Second, a control condition that participants and raters cannot identify, without which effect sizes in the psychedelic literature carry an unquantified expectancy term. Third, a regulator-accepted primary endpoint that is not a clinician-rated symptom scale, because the scales cap the reliability of everything built on them. Fourth, a payment code that reimburses assignment by measurement, since a validated marker with no code enters practice through the slowest door available. Fifth, a sponsor prepared to fund a trial sized for an interaction rather than a main effect, which is roughly four times the usual enrolment. Sixth, sensing-capable implant supply at volumes a psychiatric trial programme would need, which currently exists only because a neurological indication paid for it.
• Enables If the decisive trial succeeded, it would enable first-line treatment assignment by measurement rather than by sequence, shortening the trial-and-error period that dominates current care; it would give closed-loop psychiatric stimulation the control signal it lacks; and it would give digital phenotyping a target worth predicting. Each of those is contingent on the same single result, which is why this brief treats the programme as one question rather than four.
• Adjacent Precision Medicine supplies the pattern this field reproduces — good at predicting harm, poor at predicting benefit. Brain-Computer Interfaces shares the implanted-device supply chain and the single-subject-result problem. Cognitive Enhancement is the adjacency most often confused with this one: intervening in a healthy brain and treating a disorder have different endpoints, different regulators and different effect-size distributions.

Established “Depression is a chemical imbalance, and that is why antidepressants work.” The first clause is not supported and the second does not follow from it. A 2022 umbrella review found no consistent evidence for a serotonin abnormality as the basis of depression. Established Critics of that review correctly pointed out that it says nothing about whether the drugs work, which is a separate question answered by the trial record and answered modestly in the affirmative. Frontier Both halves of this are routinely reported as though one settled the other.

Frontier “A five-day stimulation protocol cures treatment-resistant depression.” The result behind that claim is one double-blind trial of twenty-nine participants at a single site, with a large sham-controlled remission difference at the primary timepoint. Established It is a real and important result. It is also the size at which effects in this field have historically shrunk on replication, and it does not yet distinguish individualised targeting from acceleration from total delivered dose.

Frontier “Regulators are obstructing psychedelic medicine.” The record says the evidence was contested on its merits. An advisory committee voted against effectiveness in June 2024 and the regulator asked for an additional trial in August 2024. Established The stated concerns were blinding integrity, therapy-component confounding and data-integrity questions in the trial programme, all methodological rather than ideological.

Established “Machine learning has solved outcome prediction in psychiatry.” Within-sample, often; across samples, no. A 2024 analysis of prediction models built on schizophrenia trial data found useful discrimination inside the training trial and chance-level performance in a different trial. Established Papers reporting only internal cross-validation are describing the easy half of the problem, and the field’s publication norms have not yet required the hard half.

Frontier “Pharmacogenomic testing tells you which antidepressant will work.” It predicts metabolism better than it predicts benefit. The major randomised trials of guided prescribing missed or partially missed their primary symptom endpoints while reporting positive secondaries, and independent replication has produced small, non-durable differences. Established The metabolic predictions are on firmer ground than the efficacy predictions, and commercial framing rarely separates them.

Speculative “Dimensional research frameworks have replaced diagnostic categories.” They have not, anywhere in clinical practice. Billing, trials, guidelines and regulation all still run on categorical diagnosis. Frontier The dimensional programme changed what gets funded and published and has not yet changed what gets diagnosed, and its own founding director has been among the most direct public critics of the outcome.