Cross-Source Prediction of a Visual Load Index from Vehicle Kinematic Features: A Driver-Independent Validation Study

Abstract

Highlights What are the main findings? Across held-out drivers, vehicle kinematics retained measurable information about the eye-movement-derived visual load index (VLI) proxy beyond the road-type baseline (ΔAUC = 0.042; 95% CI: 0.006–0.078). The driving-style prior increased AUC from 0.589 to 0.633, but its driver-level incremental effect remained exploratory because the confidence interval included zero. What are the implications of the main findings? By integrating within-fold proxy-label construction, source isolation, driver-grouped evaluation, and road-context controls, this study establishes a confounder-aware framework for identifying transferable information in proxy-label-based multisource sensing research. These findings position vehicle kinematics as a complementary sensing channel, rather than a replacement for eye tracking, and support its prospective evaluation in multisource driver-monitoring systems. Abstract Continuous eye tracking is difficult to deploy in production vehicles, while eye-movement-derived proxy labels are susceptible to same-source circularity and road-type confounding. We propose a driver-independent, confounder-aware cross-source framework in which the eye-movement-derived visual load index (VLI) proxy is constructed within training folds, whereas held-out prediction uses only vehicle kinematics and an optional driving-style prior. Evaluation used naturalistic data from 33 drivers, comprising 24,437 windows along a 48.4 km route, under driver-grouped cross-validation with driver-level bootstrap confidence intervals and control and increment tests. The vehicle-only model achieved an AUC of 0.589 across held-out drivers, rising to 0.633 with the style prior. After controlling for road type, vehicle features improved AUC by 0.042 (95% CI: 0.006–0.078), indicating measurable information beyond the binary road-type proxy; the style increment remained exploratory because its confidence interval included zero. Future-state prediction converged to a persistence baseline at 20 s. The VLI was not externally validated against an independent subjective or physiological criterion; accordingly, the findings are limited to cross-source prediction of an eye-movement-derived proxy. Across five architectures, this cross-source pattern was directionally consistent, supporting prospective evaluation of vehicle kinematics as a complementary sensing channel in multisource driver-monitoring systems.