Comprehensive Summary

MINDZHEIMER is a stand-alone virtual reality cognitive assessment platform for Meta Quest 3 that reframes early-stage dementia screening as a continuous behavioural-sensing problem rather than a paper-and-pencil scoring exercise. Built in Unity 6 with the XR Interaction Toolkit under a four-layer software architecture, the system administers four sequential phases — Tutorial, Object Recognition, Visuospatial fridge-placement, and Memory Recall — inside a single seated kitchen environment that eliminates locomotion as a confound for elderly users. A dual-stream data layer emits, per session, four 10 Hz trajectory CSVs capturing HMD and bilateral controller pose alongside four discrete event CSVs structured around a 23-type behavioural taxonomy that includes three novel diagnostic events (PRE_TASK_SCAN, POST_REJECTION_GAZE, OBSERVATION_SCAN_SUMMARY) operationalising orienting attention, error monitoring and encoding effort as machine-readable rows. From this dual-stream output a 32-feature schema is computed across five cognitive domains aligned to MoCA subscales — naming, visuospatial, executive, memory recall, and gaze/orienting — providing the collaborating French ML research team with a deterministic contract for feature extraction independent of the Unity codebase. System-level validation across four complete pilot sessions on two field-deployed headsets confirmed robust 9.83 ± 0.07 Hz sampling stability, 100 % SessionID continuity across all scene transitions, and zero missing files, while a healthy-versus-literature-grounded impaired feature profile contrast demonstrated that the feature space has sufficient dynamic range to express clinically meaningful differences (composite Δ ≈ 0.42). The principal contribution is therefore not a new diagnostic threshold but a complete, low-cost (≈ US$500 hardware), reproducible behavioural-sensing pipeline that bridges paper-based screening and laboratory-grade behavioural neuroscience for early Alzheimer’s detection in community settings — with convergent-validity analysis against MoCA subscores in a target cohort of 30 elderly participants and supervised ML severity classification with the French collaborators positioned as the immediate next steps.

Project Justification

The problem and why it matters. Alzheimer’s disease and related dementias affect an estimated 55 million people worldwide and are projected to reach 139 million by 2050, yet early detection in primary care still depends almost entirely on paper-and-pencil instruments — MMSE, MoCA, CDR — that were designed between 1975 and 2005. These instruments collapse a ten-minute behavioural episode into a single ordinal score and are blind to how the patient arrived at that answer: the latencies, fixations, retracings and self-corrections that distinguish a hesitant pass from a confident one. They require a trained examiner, exhibit ceiling effects in highly educated populations and floor effects in low-literacy ones, and cannot be administered at scale outside the clinic. Meanwhile, consumer VR hardware has crossed an accessibility threshold — a stand-alone Meta Quest 3 costs under US$500 and samples head and bilateral controller pose at 72 Hz with sub-centimetre accuracy — making the entire behavioural process recordable as a structured data stream. The opportunity is therefore not another VR cognitive test but the data architecture that turns any well-designed VR task into a machine-readable computational phenotype suitable for community-scale screening.

Why MINDZHEIMER is positioned to address this gap. MINDZHEIMER is built specifically around three gaps in the published VR cognitive-assessment literature: hardware accessibility (most prior systems still assume PC tethering or external eye-tracking add-ons), data output (most release a single composite score rather than the underlying behavioural stream), and validation design (few systems are explicitly engineered for convergent-validity analysis against a paper anchor). The system runs on a stand-alone Quest 3 with no external sensors, exports a complete 10 Hz dual-stream CSV output structured around a 23-event behavioural taxonomy, and is designed from the ground up for convergent-validity analysis against MoCA subscores in collaboration with a neuropsychologist. Validation across four complete pilot sessions on two field-deployed headsets has already confirmed sampling stability of 9.83 ± 0.07 Hz, 100 % SessionID continuity across scene transitions, and a 32-feature schema with sufficient dynamic range to separate healthy from literature-grounded impaired profiles. The principal deliverable is a reproducible behavioural-sensing pipeline — released independently of the Unity codebase so downstream ML teams can extract features deterministically without engine access — that bridges paper-based screening and laboratory-grade behavioural neuroscience. By doing so, the project supports the immediate research goal of supervised AD severity classification with the collaborating French ML team while creating a clinically deployable foundation for community-scale early-MCI detection at a hardware cost an order of magnitude below existing laboratory-grade alternatives.

Objectives

Phases Of Work

Estimated Budget

Project Timeline