Research prototype · NIS Almaty · 2026

Seven daily check-ins. One falsifiable forecast.

NeuroCheck fits an AR(1) model to a student's first six daily anxiety scores, then predicts day seven — and openly reports whether the prediction held. Bayes' rule, a Wilson interval and a Student prediction interval do the rest.

See the method Open the demo

0 students in the cross-sectional sample

0 consecutive days in the longitudinal series

0 held-out day used as the test point

Sample B · one student · Y₁…Y₇

95% pred. interval —

AR(1) forecast Ŷ₇ —

Actual Y₇ —

Deviation D₇ / limit T₇ —

Not met On this demonstration series D₇ = 4.071 exceeded the dynamic limit T₇ = 1.482, so the hypothesis was not confirmed. The method returned a clear, checkable answer — that is the point of it.

01 · The problem

Screening for OCD rarely reaches a classroom.

Obsessive–compulsive disorder is measurable, treatable and routinely missed in adolescence. The bottleneck is not the absence of clinical instruments — it is that no one has time to run them on every student, every week.

of the world's population lives with OCD.

WHO

Top 10

most disabling conditions worldwide.

WHO

median delay between first symptoms and treatment.

Literature

school psychologists for ~3M students in Kazakhstan.

Public estimate

Why it stays unsolved

01No routine screening. A psychologist covering several hundred students cannot run a structured interview with each of them.
02Self-report is unreliable at a single point in time. One questionnaire captures a mood, not a trajectory.
03Stigma suppresses disclosure. Adolescents under-report intrusive thoughts when the form has their name on it.
04Rare conditions break naive tests. At a 2% base rate even a good test produces far more false alarms than true ones — see section 04.

02 · Method

Four classical tools, each doing one job.

Nothing here is a black box. Every number on this page comes from a formula you can check by hand, implemented in a C++ core that writes every run to a log.

M1Bayes' rule

Turns a positive screen into a posterior probability, given sensitivity, specificity and the population base rate. This is what makes the base-rate problem visible instead of hiding it.

M2Wilson score interval

Puts honest error bars on a proportion measured in a small group. With n = 15 the interval is wide — and the site shows it that way rather than quoting the point estimate alone.

M3Student prediction interval

A range for one new observation — not for a mean. Built from the first six days (n = 6, df = 5, t = 2.571), it says where day seven should land if nothing has changed.

M4AR(1) model

A first-order autoregression fitted to that same baseline. It carries yesterday's value forward into a point forecast for today, which the held-out day then tests.

Why not a neural network?

With one participant and seven observations, any model with more than a couple of parameters would fit noise. Classical statistics is not a fallback here — at this sample size it is the only defensible choice, and it produces intervals a reviewer can audit.

How a week runs

Daily check-in

A short self-report through a Telegram bot produces one anxiety index per day.

Baseline fit

Days 1–6 give x̄, S, a prediction interval and the AR(1) coefficients.

Held-out test

Day 7 is compared against both forecasts. It never enters the fit.

Signal, not verdict

A deviation is flagged for a human to look at. The system never names a diagnosis.

03 · Evidence

Two samples, kept strictly apart.

The study answers two different questions and refuses to mix them. Sample A describes a group at one moment. Sample B tests a forecasting procedure on one person over time. Neither validates the other.

Cross-sectional

Sample A · n = 15

Above the chosen threshold8 / 15 · 53.3%

Wilson 95% interval30.1% – 75.2%

Designone questionnaire, one moment

30.1% 53.3% 75.2%

0%100%

The width of that band is the finding. With fifteen respondents the true proportion could plausibly be one in three or three in four — so this number describes these fifteen students and no one else.

Longitudinal

Sample B · n = 1 × 7

Series5, 6, 5, 7, 6, 8, 10

Baseline mean · S6.167 · 1.169

95% prediction interval[2.920; 9.413]

AR(1) estimates α, βα = 7.643 · β = −0.214

Forecast Ŷ₇ · actual Y₇5.929 · 10

D₇ vs T₇4.071 > 1.482

Day seven landed outside the prediction interval and beyond the dynamic limit, so the stated hypothesis was not confirmed on this series. A single run neither validates nor refutes the method — it demonstrates that the decision rule fires.

What the fifteen reported

Descriptive frequencies from Sample A. These are self-reported symptom frequencies, not clinical ratings.

Intrusive thoughts93.3%

53.3% often · 40% sometimes · 6.7% rarely

Repeated checking86.6%

53.3% check 1–3× · 20% 5–10× · 13.3% 10+×

Worry about locks and doors53.4%

26.7% always · 26.7% often

04 · The screening paradox

A positive screen means a 5% chance, not a 85% one.

This is the single most important number on the page, and most screening products never show it. Run 1,000 students through a test with 85% sensitivity and 70% specificity at a 2% base rate, and the arithmetic is unforgiving.

Healthy, negative Healthy, false alarm OCD, detected OCD, missed

One thousand students. At a 2% base rate, twenty of them have OCD — and you do not know which.

PPV 0

Of everyone the test flags, about one in eighteen actually has OCD. 294 false alarms against 17 true detections.

NPV 0

A negative result is usually right — but mostly because the condition is rare, not because the test is strong.

So why build it at all?

Because a 5% posterior is not a diagnosis — it is a queue. Instead of screening 1,000 students a psychologist reviews 311, and the seventeen who need help are almost certainly among them. The value is triage, and the honest framing of that is the product.

Se = 0.85 and Sp = 0.70 are deliberately optimistic scenario parameters chosen to make the arithmetic legible. They were not measured in this study. Published figures for comparable short screens sit lower — Se 0.61 / Sp 0.60 at one threshold in Piqueras et al.

05 · Limits

What NeuroCheck is not.

A screening result is not a diagnosis, and a prototype is not a validated instrument. Everything below is stated in the research paper this site is built on; repeating it here is not a disclaimer, it is the method.

✕Not a diagnosis. Crossing a threshold in Sample A, or falling outside the interval in Sample B, carries no clinical meaning on its own.
✕Not a validated scale. Se, Sp and the 25% dynamic limit are hard-coded scenario parameters, not quantities measured against a clinical reference.
✕Not generalisable. Fifteen respondents from one school describe fifteen respondents from one school.
✕Not a replacement for a specialist. The prototype proposes no treatment and makes no referral decision.
✕Not yet a secured system. In its current form the C++ core writes an unencrypted log, and the Telegram front end is not yet wired to it. Encryption, consent capture, retention limits and access control are engineering work still ahead.
✓What it is. A working computational module that turns seven self-reported numbers into an auditable, falsifiable statement — and logs every run so the statement can be checked.

Data and consent

Participation in the study was voluntary and responses were collected without identifying information. Any deployment beyond research would require informed consent before entry, minimal collection, encryption at rest, restricted access and a defined retention period. The current prototype does not yet demonstrate those guarantees, and the paper says so.

06 · Product & stack

Small, inspectable, deployable in a school.

Intake

Telegram bot (aiogram)

Compute core

C++ — statistics, PI, AR(1), Bayes

Methods

Bayes · Wilson · Student PI · AR(1)

Storage

Run log → SQLite (planned)

Position

ApproachPer-student baseline with a held-out test day, rather than a one-off cut-off score.

DataOriginal survey instrument and a seven-day pilot series collected by the author in 2026.

StatusSchool research project. No patent filed, no peer review, no clinical validation — those are milestones, not claims.


Instrument	Type	Strength	Limitation
Y-BOCS	Clinical interview	The clinical reference standard	Long; requires a trained specialist
OCI-R	Questionnaire	18 items, six subscales	Self-reported, single point in time
MOCI	Questionnaire	Short, self-administered	Measures symptoms only, not change
NeuroCheck	Computational prototype	Models an individual trajectory; every step auditable	Built on self-report; not yet validated

07 · Market

Underserved, and sized from public figures.

The numbers below are top-down estimates built on public counts of schools and students. They are a scale check, not a revenue forecast, and no school has signed anything yet.

0schools nationwide

0students aged 7–17

0school psychologists

TAM$30M

3M students × $10 / year

SAM$15M

Urban schools — roughly half

SOM$1.5M

5% of schools ≈ 350 schools

Adjacent products

Product	What it does	Where NeuroCheck differs
NOCD (US)	Teletherapy for adults with OCD	Treatment, not screening; priced far above a school budget
General wellbeing apps	Mood tracking and self-help content	No stated model, no interval, no held-out test
In-person psychologist	One-to-one assessment and support	The reference standard — but cannot cover every student weekly
NeuroCheck	Weekly per-student signal with published limits	Triage that hands a psychologist a shorter list, and shows its own error bars

08 · Business model

Three channels, one of them non-dilutive.

B2B — schools 500,000 ₸ ≈ $1,060 / year

Annual licence covering a school's students and its psychologist dashboard.

B2C — parents 5,000 ₸ ≈ $10.60 / report

An individual report for a family that wants monitoring outside school.

Grants Programme funding Non-dilutive

Digital-education and youth mental-health programmes, applied for on the same evidence base.

Year-one scenario

An illustrative target, not a booked pipeline: 100 schools and 10,000 parent reports.

50M ₸ — schools

50M ₸ — parents

100 schools × 500,000 ₸ 10,000 reports × 5,000 ₸

Total 100M ₸ / year ≈ $212,000 · at ~470 ₸/$1

09 · Roadmap

The next milestone is validation, not sales.

The paper specifies exactly what would turn this prototype into an instrument, and that study comes before any claim of accuracy.

Done

✓C++ core: sample statistics, prediction interval, AR(1), single Bayes update
✓Run logging to file for repeat inspection
✓Telegram intake prototype
✓Two-sample study written up with stated limits

→Wire the Telegram front end directly to the C++ core
→Encrypt storage; add consent capture and a retention policy
→Extend monitoring to 30 days per participant to test residual autocorrelation
→Move the threshold from hard-coded to data-selected

STAGE 1 · VALIDATION

50 clinical cases + 50 controls

Each participant independently rated with Y-BOCS or CY-BOCS, compared blind against the prototype's output.

Confusion matrix · ROC-AUC

STAGE 2 · CALIBRATION

Measured Se, Sp, PPV, NPV

Replace the scenario parameters with values measured on that sample; choose the threshold by Youden's J = Se + Sp − 1.

A defensible cut-off

STAGE 3 · PILOT

Schools, with a psychologist in the loop

Deploy only where a specialist reviews every flag, and measure whether the queue actually saves them time.

First real-world evidence

10 · Team & ask

Two people and a documented method.

Author

Kydyrbek Kozykorpesh

Nazarbayev Intellectual School of Physics and Mathematics, Almaty. Wrote the C++ core, the survey instrument and the study.

Research supervisor

Kalieva Almira

Supervised the project's design, statistical framing and interpretation limits.

Needed next: a backend engineer for the secured pipeline, and a clinical partner able to run the Y-BOCS comparison.

Use of funds — 6 months 1,500,000 ₸

≈ $3,190

Engineering — secured pipeline600,000 ₸$1,277

Validation study logistics300,000 ₸$638

Legal entity and compliance200,000 ₸$425

Hosting and equipment200,000 ₸$425

Reserve200,000 ₸$425

100 participantsValidation study size

ROC-AUCThe metric that decides whether this continues

3 schoolsSupervised pilot after validation

Get in touch

A screening signal is only useful if you can check it.

Every number on this page can be recomputed from the paper. If one of them is wrong, tell me and I will change it.

Contact the author Open the demo

koutlikefrvr4@gmail.com +7 747 426 5409