NeuroApex is a screening system, not a diagnostic one. Each level is a cheap, imperfect test that narrows who needs the next, costlier step. This page shows the maths, every test we plan to use, the papers behind them, and what we still must prove.
Start with how common a condition is (the prior). Each test result multiplies the odds by a likelihood ratio (LR). The product is the posterior: how likely the child is to have the condition given everything seen so far.
Caveat: tests of the same child are correlated (e.g. L1 CPT and L2 coding both load on attention and speed), so multiplying LRs overstates certainty. The validation plan therefore fits a joint model (logistic regression / calibrated ML) on real data rather than relying on the product rule. The calculator below is for intuition.
Estimates differ widely because of design, age band, screening tool and case definition. Each is shown with its source paper (linked in the library). Figures are the authors' own.
Sensitivity/specificity sliders are design assumptions to explore trade-offs, not measured NeuroApex performance. Measured values will come from the validation study (section 5).
Reading: with a ~7% base rate, a positive L1 alone leaves a posterior well under 50% (most L1 positives are false alarms), which is why L1 is designed for high sensitivity and is only a gate. Specificity is bought by L2 and L3, whose cost is acceptable because few children reach them. For a rare condition (ASD ≈0.2%), the same tests give a much lower PPV, so ASD relies more on parent/teacher red flags plus clinician review.
All 25 tests, using the fixed IDs from the test registry. Each card lists what it measures, how the app delivers it, which scoring features are extracted, which conditions it informs, linked evidence (abstract text is the authors' own words) and honest limitations.
Accelerometer/gyro (fidgeting), touch pressure/area, ambient audio level (no speech recorded), optional on-device camera attention metrics (no video stored), device/timing metadata. These are context signals and quality controls; they are treated as exploratory until validated.
Level 3 is delivered by clinicians, not the app. It is the pool of instruments from which a clinician chooses according to the L1/L2 profile. G = established clinical instrument or standard work-up used when indicated; R = research-only or exploratory, never used alone for diagnosis. Papers are linked where the library holds a relevant one. The G/R labels reflect general practice; the guideline-by-guideline mapping (ICMR, IAP, DSM-5, NICE etc.) is a to-do in the validation plan.
Searchable. Only the authors' own abstract text is shown, never our paraphrase of results. “PDF” opens a local open-access copy; papers without a PDF link to the open-access page (PMC) or DOI.
A screener is only useful if it is reliable, valid, fair across languages and honestly reported. This is the plan, with what exists and what is still to do.
See the sample-size paper in the library (tag STD) for the full discussion. Prevalence below is the prevalence in the validation sample (not in the population).
Rule-based scoring first, ML later and only when validated data exists. Papers in the library that report ML on screening data show the idea is feasible, but their data and populations differ from ours, so none of their accuracy numbers are transferred to NeuroApex.
Transparent z-score and percentile cut-offs per test, age-banded norms, simple risk tiers. Fully auditable.
Consented, de-identified, on-device-first pipeline. Store features (not raw audio/video). Link to clinician outcome only with consent. Data dictionary and versioning from day one.
Regularised logistic regression and gradient boosting on task features + questionnaires; calibration (Platt/isotonic), cross-validation grouped by school, reported with TRIPOD+AI. Handles correlated tests, unlike the naive LR product.
Touch-trajectory models for tracing (dysgraphia/DCD) and RT-distribution models (ex-Gaussian / drift-diffusion) for attention tasks.
Per-language/sex/school-type calibration and error audits; device-model drift monitoring; re-norming every 2–3 years.
Locked model versions, change control, human-in-the-loop review, documentation for software-as-medical-device review if claims expand. Explainable per-child “why this tier” output.
Related library papers carry the tag ML.