Figure 01 · Research design model

Detecting driver fatigue from multi‑stream naturalistic data

Design  Mixed, multilevel repeated measures
Unit of analysis  Epoch within drive within driver
Status  For review
Research design model for the driver fatigue study A left to right model showing driver level between subjects inputs and monthly organisational covariates on the left, the repeated measures drive structure with four data streams in the centre, and data fusion, modelling and outputs on the right. 01 · DRIVER LEVEL Between‑subjects Pre‑study survey · once, at T0 Age band Sex Experience (yrs) Sector / fleet Fatigue mindset Sleep habits Chronotype Health screen Baseline heart rate (resting) Fixed per driver. Groups the sample, or enters as a covariate. 02 · ORGANISATION LEVEL Time‑varying covariates Operator feed · refreshed monthly Workload index Schedule quality Rest compliance Route complexity Safety culture Policy maturity Drivers nested in operators, crossed with month. Split each index into a stable part and a monthly deviation. HOW TO READ THIS Between‑subjects Fixed per driver. Compares people. Within‑subjects Repeats. Compares a driver to themselves. Time‑varying covariate Changes month to month, not per drive. Continuous stream Sampled throughout the drive. Discrete measure Captured at fixed points only. DESIGN IN ONE LINE The two panels on the left are fixed or monthly. The centre panel repeats every drive. The three panels on the right join it together and model what predicts what. 03 · DRIVE LEVEL · WITHIN‑SUBJECTS, REPEATED The drive repeats k drives per driver over weeks to months. No cap set. Drive 1 Drive 2 Drive 3 Drive 4 ⋯ Drive k time ZOOM · INSIDE A SINGLE DRIVE Ignition on Ignition off before during after 5• Self‑report · subjective fatigue Three items on a 1‑5 fatigue scale: how fatigued before, during, after the trip. All three are completed at the end of the drive, so all three are retrospective ratings. In‑cab camera · behavioural indicators PERCLOS, blink rate and duration, eye closure, microsleep, yawning, head nod and pose, gaze dispersion, facial action units, posture shift. Visual only, no audio. Sensor logger · vehicle and context GPS latitude and longitude, speed, heading, altitude, 3‑axis accelerometer, gyroscope, harsh brake, accel and corner events, lateral sway proxy, ambient light, timestamp. Wrist wearable · physiological state Heart rate, HRV (RMSSD, SDNN), respiratory rate, skin temperature, blood oxygen, actigraphy, sleep score, time since waking. EPOCHING All streams aligned to one clock, then cut into fixed windows. The epoch is the row in your dataset. DERIVED AT DRIVE LEVEL Time of day Trip duration Distance Night driving Breaks taken Time since last sleep Cumulative hours driven Route complexity, from GPS Two nested repeated measures. Trip phase repeats inside a drive. The drive repeats across weeks. That is what makes this multilevel rather than a simple repeated measures ANOVA. groups shapes 04 · DATA FUSION One row per epoch Synchronise all streams to a common clock Cut into fixed epochs and aggregate features Attach the drive‑level self‑report ratings Join driver‑level survey data (fixed) Join operator‑month covariates (varying) Human‑code a subset to seed the labels DATA STORAGE Data will be stored and backed up securely. De‑identified, encrypted, retained per ethics approval. 05 · TWO MODELS, ONE DATASET A · Predictive model Supervised learning. Sensor and video features predict fatigue state and its precursors. Validate by holding out whole drivers, not rows, or performance will be badly overstated. B · Explanatory model Linear mixed effects. Random intercepts for driver and operator, random slope for trip phase. This is what answers the between‑subjects questions. The AI model does not. 06 · OUTPUTS What the study produces Non‑intrusive fatigue detection, validated Early warning on precursors, not just onset Which drivers are most at risk, and when Within‑driver fatigue trajectories over months Organisational levers ranked by effect Evidence base for schedule and policy change
Opposite · research design architecture