Modeling Touch Input for Users with Motor Impairments: Empirical Insights into Training Size Requirements
Radu-Daniel Vatavu, Irina Petrariu, Tudor Horomnea, Ovidiu-Ciprian Ungurean · 2026 · Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA ’26) · doi:10.1145/3772363.3798487
Summary
This CHI Extended Abstract addresses a practical question for ability-based design: how many touch observations do you actually need before you can build a user-specific model of someone's touch performance? Most touch modeling work, including Bayesian Touch and the Dual Gaussian Distribution Model, has been developed against datasets from the general population. The authors argue that touch performance for people with upper-body motor impairments varies far more between individuals and even between sessions, so a personalized model is needed, but the cost of collecting training data from disabled users is non-trivial and prior work has not quantified the minimum sample size. The team built a custom gamified Android app that records touch time (finger-down to finger-up) and offset (Euclidean distance from finger lift-off to a 15 mm target center), and collected 50 trials each from seven participants aged 45–64 with upper-body motor impairments: four with spinal cord injury at varying levels (C4–C5, T3–T4, T7), one with traumatic brain injury, one with spina bifida, and one with multiple sclerosis. WHODAS 2.0 scores ranged from 23.0 to 68.7. They fit per-user Gaussian models for time and offset under varying training-set sizes (N = 2, 4, 8, 16, 24, 32, 48), repeating the random sampling 1000 times, and evaluated each via three orthogonal checks: predictive coverage of the 95% interval, predictive bias of z-scores against zero, and average predictive log-likelihood.
Key findings
Inter-participant variability was extreme. Mean touch time ranged from 58 ms (P1, spinal cord injury C4–C5) to 1563 ms (P6, also spinal cord injury C4–C5), a 27× spread, and mean offset ranged from 2.5 mm to 10.5 mm. This alone refutes any one-size-fits-all touch target heuristic for disabled users and confirms Findlater et al.'s warning against predicting performance from diagnosis or age alone. Models trained on N=2 or N=4 observations were unusable: coverage dropped as low as 68% (against a 95% target), z-score bias drifted to 1.8–2.1, and log-likelihood ran to −2045. From N=8 onwards, log-likelihood plateaued for most participants and coverage and bias stabilized close to nominal. From N=24 to N=48 several participants (P1, P2 for time; P5, P6) showed signs of overfitting, with coverage drifting either above or below 95%. The authors conclude that 8–24 user-specific touch observations is the sweet spot, and they translate this into a four-step Observe–Model–Revise–Share procedure aligned with Wobbrock et al.'s ability-based design principles, with concrete examples (resize menu options based on the offset model; personalize long-press thresholds based on the time model).
Relevance
For practitioners shipping adaptive interfaces, the headline finding is unusually actionable: you can build a useful user-specific touch model from 8 to 24 taps with nothing more than basic on-device math, no ML framework required. That makes ability-based personalization tractable for any team that has been put off by the cost of model training or the data-collection burden on disabled users. The framing also pushes back against the assumption that you need deep learning to do this well; for touch time and offset, a Gaussian with Bessel-corrected variance and a sliding window outperforms heavier approaches in deployability. Limitations are real: N=7 is small, the Gaussian assumption breaks down for highly variable users (P3 and P6 needed more data to plateau), and the work covers only target acquisition, not gestures, dwell, or multi-touch. Cross-application sharing of learned profiles is proposed but not implemented, and raises its own privacy questions that the paper does not address.
Tags: touch input · motor impairment · ability-based design · adaptive interfaces · touchscreen accessibility · user modeling · personalization · mobile accessibility
Standards referenced: WHODAS 2.0 (WHO Disability Assessment Schedule) · Material Design accessibility guidelines · Microsoft touch target guidelines