Bulbitech experts Ivania Patry and Ping Zhao discuss the real-world value, limitations, and future role of AI-driven eye-tracking in ophthalmology and neuro-ophthalmic care.
What are the common misconceptions about AI in eye-brain assessment?
A common misconception is that AI directly "sees" pathology. In reality, it identifies statistical patterns and is highly dependent on the quality and labeling of its training data. Another misconception is that AI provides definitive diagnoses — most tools offer probabilistic risk scores requiring clinical interpretation.
There is also a false belief that AI eliminates variability. In practice, performance may fluctuate with multiple factors including camera hardware, lighting, or patient demographics. The perceived "black box" nature of deep learning models also challenges expectations of transparency in clinical medicine. More data does not automatically improve accuracy; biased or poorly curated datasets may amplify error.
Finally, while AI is often associated with efficiency gains, early implementation frequently introduces additional workflow steps. AI is best understood as a clinical decision-support tool rather than a replacement for clinical reasoning.
Where can objective eye-tracking data provide real clinical value, and where might it fall short?
Objective video-oculography provides growing value by quantifying subtle metrics, such as saccadic latency, smooth pursuit gain and fixation stability, impossible to measure precisely with the naked eye. It also standardizes longitudinal data, reducing inter-examiner variability.
This makes video-oculography a useful add-on tool for both screening and diagnosis, but also for prognostic evaluation, longitudinal patient follow-up, and assessment of treatment effects. However, video-oculography should not be regarded as a standalone diagnostic test. Oculomotor abnormalities are often non-specific, and factors such as fatigue, anxiety, or medication effects may influence results.
Technical limitations, including poor calibration, ptosis, anatomical variation, or limited patient cooperation, can reduce data quality. Cost-effectiveness has yet to be established in many healthcare settings. Oculomotor measures are constrained by visual perception and motor execution, whereas pupillary and reflex-based measures can remain robust in conditions dominated by sensory loss or motor impairment.
What should ophthalmologists consider before adopting AI-assisted assessment tools?
Clinical validation is central: Has the tool been evaluated in populations and disease groups relevant to your practice? Does its performance profile align with its intended use, for example, prioritizing sensitivity in screening settings or specificity in confirmatory contexts?
Real-world performance may differ from results reported in controlled studies due to variations in hardware, workflow, and patient demographics. Explainability also matters: does the system provide interpretable outputs that support clinical decision-making?
Practical integration should be assessed. Compatibility with electronic medical records, interoperability, and data flow are key considerations. Algorithmic performance may drift over time and require revalidation, and this process should be transparent.
Financial and regulatory factors are equally important, including reimbursement pathways, data protection requirements (such as GDPR, or HIPAA in the US), and liability considerations. Adoption should be guided by demonstrable clinical value rather than technological novelty.
How can AI complement traditional assessments without replacing clinician expertise?
AI complements clinicians by acting as a high-precision measurement tool, while the clinician serves as integrator and communicator. AI excels at repetitive, quantitative tasks, measuring millisecond-level saccade latencies or analyzing vast datasets for subtle trends often missed by human observation.
In some settings, AI systems are being explored for triage or prioritization of higher-risk cases, although their impact on efficiency varies. Importantly, AI systems lack contextual understanding. They do not incorporate subjective symptom history, complex comorbidities, psychosocial factors, or patient experience.
A “human-in-the-loop” model remains essential. AI may provide structured data and probabilistic outputs, but the clinician interprets findings, integrates them within the broader clinical picture, and determines management. Clinical judgement and communication remain central to patient care.
What are the practical challenges clinicians face when integrating AI-driven eye-tracking into their work?
Workflow, interoperability, and cost are significant challenges. Incorporating hardware into cramped exam rooms and allocating calibration time can disrupt clinic flow. Patient factors, including the inability to follow instructions (e.g., pediatric or cognitive impairment cases) or physical limitations (e.g. ptosis), may lead to delays and data quality degradation.
From an information technology perspective, integrating AI outputs into EMRs remains difficult, often trapping data in silos. Data governance raises additional questions, including ownership and secondary use of patient data for algorithm refinement.
Clinicians and staff require training to ensure appropriate data acquisition and interpretation of probabilistic outputs. Reimbursement pathways are not yet clearly defined in many regions, which may limit adoption, particularly in smaller practices or publicly funded systems.
Are there specific ocular or neurological conditions where AI eye-tracking has shown particular promise or limitations?
Quantitative eye movement analysis may be particularly informative in conditions where assessment relies heavily on clinical examination or subjective reporting. This includes certain neurodegenerative disorders, concussion, and selected neurodevelopmental or psychiatric conditions, although evidence remains evolving.
It may also support pre- and post-operative assessment in some surgical contexts, such as strabismus or ptosis, by providing objective longitudinal comparison.
No single oculomotor abnormality is pathognomonic for a specific disease. However, particular patterns, such as vertical gaze palsy or specific forms of nystagmus, enabling the mapping of functional deficits to anatomical substrates. Interpretation must always be integrated with broader neurological and ophthalmic findings.
Looking ahead, how do you see AI-driven eye-tracking shaping ophthalmology and neuro-ophthalmic care over the next 3–5 years?
Over the past few years, several AI-based medical devices have received clinical approval in Europe, predominantly in imaging. A number of video-oculography devices have also been approved, although these do not currently incorporate AI-driven analysis. The integration of AI into video-oculography diagnostics is now beginning to emerge, with the aim of moving these technologies beyond research settings towards clinically validated tools capable of delivering objective, quantitative assessments.
This development is likely to be particularly relevant in neurology and ophthalmology, and potentially in other specialties where systemic diseases, such as diabetes or thyroid disorders, may have ocular manifestations.
A key challenge will be shifting from single time-point assessments and relatively small study cohorts to large-scale, longitudinal evaluation in real-world populations. The development of broad, population-representative normative datasets that account for demographic variation should help refine diagnostic sensitivity and specificity, and increase clinical confidence. Studies across different stages of disease may also support earlier detection, improved prognostic assessment, and more accurate stage stratification. Integration with complementary imaging and diagnostic modalities could further enable multimodal risk stratification and clinical decision support.
Ivania Patry, Chief Medical Officer, Bulbitech AS, is a neurologist with extensive experience as Principal Investigator in clinical trials and regulatory affairs leadership.
Ping Zhao, Senior AI Engineer, Bulbitech AS, is a software engineer with over 12 years of experience in machine learning algorithm development, image processing, and real-time system optimization.
