Reinforcement-Based Head-Scan Training for Drivers with Homonymous Visual Field Loss Improves Blind-Side Detection

Abstract

Objective:

To evaluate whether a reinforcement-based head-scanning training program improves blind-side hazard detection in drivers with homonymous visual field loss.

Design:

Single-arm non-randomized clinical trial with 1-month follow-up.

Setting:

Research laboratory.

Participants:

Inclusion criteria included ≥3 months post-onset hemianopia or quadrantanopia, no neglect or major cognitive decline, and current or former driving experience. Sixteen of 23 volunteers met criteria and completed the study (median age = 52.5 years; 56% left-sided field loss).

Interventions:

Participants completed three 1-hour sessions of a reinforcement-based head-scanning training program in a driving simulator. An auditory reminder cue served as immediate, error-contingent feedback, sounding from a loudspeaker on the participant’s blind side if they failed to make a large, early head scan to their blind side before an intersection.

Main Outcomes Measures:

The primary outcomes were rates of proactive blind-side scanning (the proportion of intersections with an early large head scan) and blind-side hazard detection (the proportion of hazard detected) assessed without reminder cues at pre-training, 1 week post-training, and 1 month post-training. All hypotheses were prespecified.

Results:

Proactive blind-side head scanning increased significantly from 17.5% (95% CI: [8.9%, 26.0%]) at pre-training to 83.6% (95% CI: [76.5%, 90.7%]) at 1 week post-training and 72.6% (95% CI: [63.4%, 81.9%]) at 1 month post-training. Likewise, blind-side hazard detection improved significantly from 55.6% (95% CI: [41.3%, 69.9%]) to 85.6% (95% CI: [76.3%, 95.0%]) at post-1-week and 82.5% (95% CI: [70.1%, 94.9%]) at post-1-month. Improvements were primarily due to fewer scanning-related errors.

Conclusions:

Three sessions of reinforcement-based head-scanning training significantly improved blind-side head scanning and hazard detection in drivers with homonymous visual field loss, with benefits lasting at least one month. These results support the promise of this rehabilitation approach for post-stroke driving and warrant further evaluation in a randomized controlled trial.

About 1 in 4 adults will have a stroke at age 25 or older.¹ Up to 60% of stroke survivors develop permanent visual field loss,^2,3 most commonly homonymous visual field loss (HVFL), caused by damage to the post-chiasmal visual pathways. It results in loss of vision on the same side of the visual field in both eyes, affecting one half (hemianopia) or one quarter (quadrantanopia) of the visual field depending on lesion location. HVFL can significantly impair daily functioning, reduce independence in mobility and driving, limit employment, and increase depression risk.^4–6 Resuming driving is a major rehabilitation goal for many stroke survivors with HVFL. Driving with HVFL is a realistic possibility if the person meets the minimum visual field requirements for driving in the jurisdiction where they reside⁷ and does not have any other contraindications, such as impaired cognitive capacity or significant spatial neglect.

Although individuals with HVFL may be permitted to drive, their risk of at-fault collisions is more than twice that of drivers without vision loss.⁸ To compensate for the field loss, drivers with HVFL need to scan toward the affected side (blind side) to bring hazards into view. At intersections, they may need to scan up to 90°, requiring a large head movement to their blind side.⁹ However, many individuals with HVFL do not spontaneously adapt their scanning adequately to traffic demands.¹⁰ Real-world^11,12 and simulator^9,13–15 studies show that they often fail to scan sufficiently toward the blind side, even at high-risk intersections, and may remain unaware of these lapses unless there is a consequence such as a near collision or external feedback occurs.¹⁶ Although scanning training programs for patients with HVFL have been evaluated for reading, visual search and walking,^17–22 there is a paucity of research on scanning training for driving,²² especially training to make large head scans at intersections. The current study addresses this important gap.

Prior driving research showed that early, large head scans toward the blind side during intersection approach, particularly those exceeding 20°, were associated with improved hazard detection.¹³ In a proof-of-concept driving simulator study, real-time auditory reminder cues, delivered when such scans were absent, further promoted proactive scanning with HVFL participants spontaneously adopting early large scans to avoid the cues, improving blind-side detection.²³ However, that study involved only a single session with real-time cues, leaving it unclear whether the proactive scanning behavior would persist once cues were removed or whether the method could be scaled as a training tool. To address this, we developed a reinforcement-based head-scan training intervention and demonstrated its efficacy in a small pilot study.²⁴ Using auditory reminder cues as negative reinforcement, participants were instructed to make early, large head scans toward the blind side when approaching intersections; a “beep” sounded from that side if they failed, providing immediate error-contingent feedback. The cue served as a consequence that provided the feedback needed for learning, reinforcing proactive blind-side scanning.

Here, we report a non-randomized clinical trial to evaluate the effects of our novel intervention on proactive blind-side scanning and hazard detection. Participants with HVFL completed six sessions: one pre-training evaluation, three weekly training sessions, and two follow-up evaluations at 1 week and 1 month after training. We hypothesized that the head-scan training paradigm would produce improvements in blind-side scanning evident even after the reminder cues were removed. We further hypothesized that the improvements in blind-side scanning would result in improved blind-side hazard detection.

Methods

Study Design

This single-arm, non-randomized clinical trial was conducted at the Schepens Eye Research Institute from June 2024 to May 2025. The study adhered to the Declaration of Helsinki and Good Clinical Practice and followed the Consolidated Standards of Reporting Trials (CONSORT) reporting guidelines. Procedures were approved by the Mass General Brigham Institutional Review Board and written informed consent was obtained at the start of the first visit. The trial was registered on ClinicalTrials.gov (NCT06136169) on 18 November 2023. The trial protocol, CONSORT checklist and participant flow diagram are provided in Supplements 1, 2 and 3, respectively.

Participants

Participants were recruited from outpatient clinics. Inclusion criteria were chronic HVFL (hemianopia or quadrantanopia; Goldmann V4e) with at least three months post-onset to minimize spontaneous recovery, binocular visual acuity of LogMAR (Logarithm of the Minimum Angle of Resolution) ≤ 0.3, current or prior driving experience, and no visual neglect. Neglect was screened using the Gauthier Bells Test²⁵ (left-right asymmetry in missed bells ≥ 4) or Schenkenberg Line Bisection Test²⁶ (left-right asymmetry in missed lines ≥ 2 or > 11% deviation from the middle). All participants completed the Montreal Cognitive Assessment (MoCA) test.²⁷

Driving Simulator

All sessions were conducted in a high-fidelity driving simulator (Model LE-1500, FAAC Corp.) with a 225° horizontal and 32° vertical field of view on the center screen (Figure 1). Guided by auditory GPS instructions, participants drove along predetermined routes (each about 7 minutes) through a city with many intersections, oncoming and cross traffic. Head movements were continuously tracked using a remote six-camera system (SmartEye Pro 6.1; Smart Eye AB) with a 180° horizontal range (±90°).

Figure 1. Driving simulator setup with integrated head and eye tracking.

Six infrared tracking cameras (blue circles) are positioned around the driver’s seat: two on the left, two in the center, and two on the right. In this image, the left-side cameras are obscured by the driver and seat. The cameras capture head and eye movements across a 180° horizontal field. Auditory reminder cues are delivered through loudspeakers mounted above on the left and right sides (yellow circles).

Intervention

Participants completed three reinforcement training sessions with head scan reminders, each about an hour, one week apart. Each session involved driving three city routes containing 24 – 25 intersections. The first training session began with an introductory video explaining HVFL and the importance of compensatory head movements, particularly at intersections. In addition to standard “left-right-left” scanning at intersections, participants were instructed to make an early, large head scan toward the blind side before reaching the intersection. They were told that this extra head scan would help to compensate for the field loss by providing an early preview of whether there were any hazards on that side. Auditory reminder cues (a 1-second beep from a loudspeaker at about 45° on the blind side) were used to reinforce this behavior. Cues were delivered when the participant’s vehicle was 20 m from the intersection. In the first training drive participants received a reminder cue at every intersection to establish the habit of making an early large head scan. In all subsequent training drives, head position was monitored in real time, and a reminder cue was delivered only if no early, large blind-side scan (≥20°) occurred between 50 and 20 m before the intersection.^13,23,24 Participants received end-of-session feedback summarizing the proportion of intersections where the cue was triggered.

Evaluation

Each evaluation session involved driving two city routes with 26 – 27 intersections. No reminder cues or performance feedback were provided. Motorcycle hazards appeared on the cross street at 10 intersections in each drive (five from the left and five from the right, pseudorandomized). Each hazard emerged at 55°–67° eccentricity, approached the intersection at 40 mph, did not slow down, and disappeared just before entering the driver’s path (eFigure 2, Supplement 3). Participants pressed the horn when they detected a motorcycle.

Outcome Measures

Head movements were analyzed offline using a validated algorithm^28–30 to detect head scans (see eTable 1, Supplement 3 for head scan thresholds). The primary scanning outcome was the rate of proactive scanning, defined as the proportion of intersections with at least one early (50 – 20 m before the intersection), large (≥ 20°) head scan toward the blind side. Secondary scanning outcomes were the number of head scans per intersection and the magnitude of head scans, both toward the blind and seeing sides, during the 100 m approach to each intersection.^9,29,30

The primary hazard detection outcome was detection rate, defined as the proportion of motorcycle hazards identified with a horn press; missed blind-side detections were further classified as scanning errors (no or insufficient scan) or other errors (adequate scan but no response). The secondary hazard detection outcome was response time, defined as the latency from motorcycle appearance to horn press.

Statistical Analysis

The target sample size was 16 participants, based on prior²³ and pilot²⁴ data showing large cue-related improvements in blind-side scanning (Cohen’s d ≈ 0.8) in metrics analogous to this trial’s primary outcomes. This sample provided approximately 80% power to detect within-subject differences (paired t test; α = .05).

The primary analyses examined (1) the proportion of intersections with an early, large blind-side scan and (2) the proportion of detected blind-side hazards before versus after training. These outcomes were analyzed using generalized linear mixed models (GLMMs) with a fixed effect for session (pre-training as reference). The hazard detection model additionally included motorcycle approach side (blind or seeing) as a fixed effect. Head scanning data were analyzed at the intersection level; hazard detection data were analyzed at the event level.

The secondary analyses evaluated scan count, scan magnitude, and hazard response time. Intersections in which participants received a reminder cue during training were excluded from the head scan count and magnitude analyses. Scan count (the number of scans per intersection) was analyzed using a GLMM with fixed effects for session and head scan direction. Scan magnitude (median scan magnitude per intersection) and response time were analyzed using linear mixed models (LMMs); scan magnitude and response time were log-transformed prior to analysis. The LMM for scan magnitude included fixed effects for session and head scan direction, whereas the LMM for response time included fixed effects for session and motorcycle approach side.

All models were constructed in RStudio (R version 4.3.2) using the lme4 package. All GLMMs and LMMs included random effects for participant and intersection. Prespecified contrasts compared each visit with pre-training within each side. Results are reported as raw (observed) subject-level means with 95% confidence intervals; inferential statistics (Z-ratios, T-ratios, and p-values) are from model-based contrasts. Statistical tests were two-sided, and results with P < .05 were considered statistically significant.

Results

Participant Characteristics

Twenty-three individuals with HVFL were screened for eligibility; seven did not meet study criteria (eTable 2, Supplement 3). Sixteen met the inclusion criteria, were assigned to the intervention, and completed the study. Sample characteristics are shown in Table 1: median age, 52.5 years (range, 17–76); 7 participants (44%) were female; and 9 (56%) had left-sided visual field loss.

Table 1.

Demographic and clinical characteristics of participants (N = 16).

Characteristic	Value
Age, median (range), years	52.5 (17 – 75)
Gender, No. (%)	7 (44) females, 9 (56) males
Race, No. (%)	15 (94) White, 1 (6) Asian
Side of visual field loss, No. (%)	9 (56) left, 7 (44) right
Type of visual field loss, No. (%)	10 (63) hemianopia, 6 (38) quadrantanopia
Cause of visual field loss, No. (%)
Stroke	9 (56)
Surgery	4 (25)
Trauma	1 (6)
Other	2 (13)
Time since onset, median (range), years	5 (1 – 25)
Current driver, No. (%)	5 (31)
Binocular visual acuity, median (range), LogMAR21	−0.25 (−0.3 – 0.05)
Montreal cognitive assessment score, median (range)	27 (21 – 30)

¹Logarithm of the Minimum Angle of Resolution

Head Scanning

Proactive blind-side scanning.

Training significantly improved proactive blind-side scanning (Figure 2). Before training, participants made an early, large head scan to their blind side at only 17.5% (95% CI: [8.9%, 26.0%]) of intersections. This increased significantly to 76.9% (95% CI: [69.5%, 84.3%]) in Training 1 and reached near-ceiling levels in Training 2 (M = 94.3%, 95% CI: [91.8%, 96.7%]) and Training 3 (M = 95.3%; CI: [92.3%, 98.3%]). These gains persisted without reminder cues: proactive scanning remained high at 1-week post-training (M = 83.6%, 95% CI: [76.5%, 90.7%]) and 1-month post-training (M = 72.6%, 95% CI: [63.4%, 81.9%]). Planned contrasts from the GLMM comparing each visit with pre-training showed that the rate of proactive blind-side scanning was significantly higher at all sessions relative to pre-training (all Ps < .0001).

Figure 2. Proactive blind-side scanning across study sessions.

Proportion of intersections where participants made an early, large head scan (≥ 20° toward the blind side between 50 – 20 meters before the intersection) across study sessions. Solid, filled circles show group means with between-subjects 95% confidence interval; individual participant data points (translucent red circles) are overlaid. The shaded region indicates the training phase (T1 – T3), during which auditory reminder cues were provided for missed or insufficiently large blind-side scans. No reminder cues were given during the pre- or post-training evaluation sessions.

Head scan count.

Participants made an average of 3.57 scans per intersection (95% CI: [2.73, 4.41]) before training (data pooled across blind and seeing sides), which increased significantly to 5.40 (95% CI: [4.74, 6.06]) one week after training and 5.00 (95% CI: [4.29, 5.72]) one month after training (main effect of sessions: Ps < .005). Participants also made more scans to their blind (M = 3.29, 95% CI: [2.91, 3.68]) than seeing side (M = 1.97, 95% CI: [1.69, 2.25]) (main effect of scan direction: P < .0001). Planned contrasts showed that participants made significantly more scans per intersection to the blind side at both post-1-week and post-1-month compared to pre-training (Ps < 0.0001; Figure 3A and Table 2). On the seeing side, scan count did not change significantly after training (Ps > 0.45; Figure 3A and Table 2).

Figure 3. Head scanning behavior during intersection approach across study sessions.

(A) Head scan count (number of head scans per intersection) and (B) head scan magnitude (in degrees) from 100 to 0 meters before the intersection. Scans are grouped by direction: blind side (red circles) and seeing side (blue triangles). Solid, filled shapes represent group means with between-subjects 95% confidence interval, and individual participant data (translucent shapes) are overlaid.

Table 2. Results summary for head scan count and magnitude.

Means are raw values, and 95% confidence intervals represent between-subject variability. Test statistics (Z-ratios and T-ratios) and P values were obtained from planned contrasts comparing each visit with the pre-evaluation session, conducted separately for blind and seeing sides using Dunnett-adjusted contrasts of estimated marginal means from the GLMM and LMM.

	Pre-Training	Post-1-Week		Post-1-Month
Measure	Mean [95% CI]	Mean [95% CI]	Test stats [P value]	Mean [95% CI]	Test stats [P value]
Head scan count (mean number of scans per intersection)
Blind side	2.21 [1.75, 2.67]	3.53 [3.08, 3.98]	13.76 [<.0001]	3.16 [2.68, 3.65]	10.08 [<.0001]
Seeing side	1.95 [1.58, 2.32]	2.05 [1.75, 2.35]	0.37 [.978]	2.11 [1.77, 2.44]	1.44 [.457]
Head scan magnitude (°)
Blind side	23.9 [20.3, 27.5]	27.5 [23.8, 31.2]	9.39 [<.0001]	28.0 [23.9, 32.0]	9.60 [<.0001]
Seeing side	24.0 [20.8, 27.2]	28.1 [24.1, 32.0]	8.09 [<.0001]	27.2 [23.3, 31.1]	6.71 [<.0001]

Head scan magnitude.

Compared with pre-training (M = 23.9°, 95% CI: [21.0°, 26.9°]), participants made larger head scans at post-1-week (M = 27.8°, 95% CI: [24.2°, 31.4°]) and post-1-month (M = 27.6°, 95% CI: [23.8°, 31.3°]) evaluations (data pooled across blind and seeing sides) (main effect of sessions: Ps < .02). There was no significant difference between the magnitude of the scans made to the blind side and seeing side (blind side: M = 28.6°, 95% CI: [24.4°, 32.8°]), seeing side: M = 29.2°, 95% CI: [25.8°, 32.6°], main effect of scan direction: P = .423). Planned contrasts showed that scans to both blind and seeing sides were significantly larger at both post-1-week and post-1-month evaluations compared to pre-training (Ps < 0.0001; Figure 3B and Table 2).

Hazard Detection

Detection rate.

Blind-side motorcycle detection improved significantly after training (Figure 4 and Table 3), increasing from mean detection rate of 55.6% (95% CI: [41.3%, 69.9%]) at pre-training to 85.6% (95% CI: [76.3%, 95.0%]) at post-1-week and 82.5% (95% CI: [70.1%, 94.9%]) at post-1-month (Ps compared to pre-training < 0.0001). In contrast, detection on the seeing side remained high and stable across sessions (pre-training: 93.1% (95% CI: [87.4%, 98.9%]), post-1-week: 95.6% (95% CI: [90.0%, 100.0%]), post-1-month: 93.1% (95% CI: [88.5%, 97.8%]), Ps > 0.505; Table 3).

Figure 4. Hazard detection before and after training.

Motorcycle detection rate at pre-training, 1 week post-training, and 1 month post-training, shown separately for motorcycle approach from the blind side (red circles) and seeing side (blue triangles). Solid, filled shapes represent group means with between-subjects 95% confidence interval; individual values (translucent shapes) are overlaid.

Table 3. Results summary for hazard detection rate and response time.

Means are raw values, and 95% confidence intervals represent between-subject variability. Test statistics (Z-ratios and T-ratios) and P values were obtained from planned contrasts comparing each visit with the pre-evaluation session, conducted separately for blind and seeing sides using Dunnett-adjusted contrasts of estimated marginal means from the GLMM and LMM.

	Pre-Training	Post-1-Week		Post-1-Month
Measure	Mean [95% CI]	Mean [95% CI]	Test stats [P value]	Mean [95% CI]	Test stats [P value]
Detection rate (%)
Blind side	55.6 [41.3, 69.9]	85.6 [76.3, 95.0]	6.13 [<.0001]	82.5 [70.1, 94.9]	5.51 [<.0001]
Seeing side	93.1 [87.4, 98.9]	95.6 [90.0, 100.0]	1.0 [0.505]	93.1 [88.5, 97.8]	0.0 [1.0]
Response time (s)
Blind side	2.72 [2.41, 3.04]	1.85 [1.57, 2.12]	9.10 [<.0001]	2.13 [1.87, 2.40]	6.17 [<.0001]
Seeing side	1.83 [1.66, 2.00]	1.90 [1.63, 2.16]	0.18 [0.969]	1.91 [1.68, 2.13]	0.24 [0.947]

Reasons for missed blind-side detections.

Reductions in missed blind-side detections were primarily driven by reductions in scanning-related errors (Figure 5). The most common error at pre-training was making a scan that was not large enough toward the blind side, which dropped from 28.8% pre-training to 8.1% at post-1-week and 10.6% at post-1-month. Missed detections due to failure to scan to the blind side declined from 10.6% pre-training to 3.8% at post-1-week and 4.4% at post-1-month, respectively. Other infrequent errors (< 5%) involved participants looking at the hazard but not responding, possibly due to momentary lapses in attention or response failures.

Figure 5. Reasons for missed blind-side hazard detection before and after training.

Proportion of blind-side motorcycle hazards classified by detection outcome across evaluation sessions. Stacked bars show the relative frequency of four categories: detected (green), and three categories of misses, fixated but no response (orange), scanned with insufficient head magnitude (light red), and no blind-side scan (dark red).

Response time.

Participants also responded to blind-side motorcycles significantly faster after training (Figure 6, Table 3). Before training, they took an average of 2.72s (95% CI: [2.41s, 3.04s]) to respond to motorcycles approaching from their blind side. This reduced to 1.85s (95% CI: [1.57s, 2.12s]) at the post-1-week evaluation and 2.13s (95% CI: [1.87s, 2.40s]) at the post-1-month evaluation (Ps compared to pre-training < 0.0001). On the other hand, seeing-side response times varied little across sessions (Ps > 0.947; Table 3).

Figure 6. Response time to detected motorcycle hazards before and after training.

Mean response time (in seconds) for hazards approaching from the blind side (red circles) and seeing side (blue triangles) during pre-training, 1 week post-training, and 1 month post-training. Solid, filled shapes show group means with between-subjects 95% confidence interval; individual data points (translucent shapes) are overlaid.

Discussion

This clinical trial shows that three one-hour sessions of reinforcement-based head-scan training rapidly improved proactive blind-side scanning in individuals with HVFL, despite heterogeneous performance before training. Such consistent acquisition is notable for a clinical population. Training effects translated into functional gains relevant to driving safety: blind-side hazard detection increased by about 30%, with improvements persisting for at least one month, and seeing-side performance remained stable. Participants also detected blind-side motorcycle hazards earlier. These gains were driven mainly by fewer scanning-related errors, specifically absent or insufficient blind-side scans, and are unlikely due to test familiarity because seeing-side performance was unchanged and prior simulator studies found no improvement from repeated testing alone.³¹

These results build on the prior proof-of-concept study²³ showing that real-time auditory cues can prompt proactive blind-side scanning. By applying reinforcement learning principles, the present trial implemented auditory reminder-cue training that produced rapid behavioral change lasting at least one month. To our knowledge, this is the first clinical trial to demonstrate that driving-specific head-scan training improves blind-side scanning and hazard detection at intersections. The findings support targeted compensatory scanning programs for driving rehabilitation in individuals with HVFL.

Several factors likely contributed to the effectiveness of the intervention. Training combined explicit instruction, hands-on practice in a realistic simulator, and immediate, error-contingent feedback, allowing participants to learn the desired behavior and correct mistakes in real time. This design supports both declarative learning (facts and concepts that can be verbalized) and procedural learning (sequences of steps that can be enacted), which together are central to skill acquisition.³² The new behavior of making an early large blind-side head scan was also stacked onto existing habits, such as familiar left-right-left scanning routines at intersections, which likely facilitated learning.³³ The execution of the behavior was further reinforced through negative reinforcement: auditory cues were delivered only when scans were absent or insufficient, motivating participants to “beat the beeps.”³⁴ Training was spaced across multiple sessions, which is consistent with a large body of work in cognitive and educational psychology showing that spacing practice over time leads to stronger long-term learning than massed practice.³⁵ Finally, training focused on a transferable skill (proactive blind-side scanning), supporting its generalization to untrained contexts, such as motorcycle detection in novel scenarios.

In the United States, rehabilitation for drivers with HVFL largely focuses on scanning training, vehicle adaptations (e.g., panoramic rear-view mirrors) and behind-the-wheel training by certified driving rehabilitation specialists.^16,36 However, a nationwide survey found that only 11 of 57 drivers with HVFL reported receiving driving-specific rehabilitation.³⁶ Peripheral prism glasses^37,38 that expand the field of view are another rehabilitation option and have been shown to improve blind-side detection in early on-road and driving simulator studies.^11,39 Although scanning training is often considered the primary type of visual rehabilitation for people with HVFL, the specific characteristics of effective scanning training for driving have not been well defined.²² The present study is the first step in addressing this knowledge gap.

Study Limitations

First, single-arm non-randomized clinical trials may have greater risk for bias in the results than randomized controlled trials. However, we suggest that the risk of bias was low because the primary outcomes were based on objective measures of performance in evaluation drives without feedback about scanning performance or detection performance. There was no feedback from research personnel, no scanning reminder cues, and motorcycle hazards disappeared before entering the participant’s travel path, so participants were likely largely unaware of insufficient scanning or missed detections in evaluation drives. The absence of a control condition was a limitation; improvements in scanning could, in principle, reflect repeated exposure to driving in the simulator, though this seems unlikely based on prior studies.³¹ Future studies should include a control condition, such as comparing training with versus without auditory reminder cues, to determine whether the cues themselves (and “beating the beeps”) are the key to the observed improvements in scanning behavior measured in the current study.

Second, our sample was heterogenous, including current and former drivers with hemianopia or quadrantanopia, mostly recruited from an outpatient population attending a local eye hospital. They may have been more motivated to resume driving than the broader post-stroke population. However, this reflects the intended clinical population - individuals with HVFL who are likely to seek driving rehabilitation in outpatient settings. Although all participants in the current study acquired proactive blind-side scanning behavior quickly, there was variability in pre-training performance and post-training retention. Our study was not powered for sub-group analyses. Future studies should recruit larger samples to enable subgroup analyses and systematically identify which individuals benefit most.

Third, this first clinical trial of the reinforcement-based head-scan training focused on short-term follow up (to one month) with evaluations in the driving simulator. Future research should examine long-term maintenance and evaluate strategies such as periodic booster sessions, particularly for subgroups that may benefit most. It will also be important to assess whether training effects generalize to other driving scenarios such as non-intersection locations, real-world driving,¹² and other mobility challenges that would benefit from similar compensatory behaviors, such as scanning before crossing busy streets when walking.⁴⁰

Conclusion

This clinical trial demonstrates that reinforcement-based head-scan training using real-time behavior-contingent feedback produces rapid improvements in compensatory scanning at intersections among individuals with HVFL. Three one-hour sessions were sufficient to establish proactive blind-side scanning, with effects translating into better detection of blind-side hazards and persisting for at least one month. These results demonstrate feasibility and clinical promise for this rehabilitation approach and suggest its potential to improve driving-related compensatory visual scanning of stroke survivors with HVFL.

Abbreviations:

CONSORT

Consolidated Standards of Reporting Trials

GLMM

generalized linear mixed model

HVFL

homonymous visual field loss

LogMAR

logarithm of the minimum angle of resolution

LMM

linear mixed model