The Role of Artificial Intelligence in Dry Eye Diagnosis

Dry eye disease (DED) remains one of the most common and frustrating conditions encountered by eyecare professionals, particularly within optometry. It is a leading cause of ocular discomfort, blurred vision, and fluctuating visual performance, often interfering with reading, driving, computer use, and contact lens wear.¹

The chronic nature of DED, coupled with its impact on quality of life and workplace productivity, underscores the need for accurate and timely diagnosis and treatment. Despite decades of research, diagnosing DED often feels as much like an art as a science, with clinical signs that can be subtle, variable, or entirely absent despite significant patient discomfort.

Artificial intelligence (AI) has emerged as a promising tool to aid in addressing these challenges. By leveraging advanced imaging, pattern recognition, and data analysis, AI offers the potential to improve diagnostic accuracy, reduce variability, and streamline clinical workflow.²

The global market for AI in eyecare is projected to grow substantially over the next decade, driven by demand for early disease detection and precision care, making DED an ideal candidate for AI-enhanced evaluation.³ As adoption accelerates, AI is poised to change how eyecare professionals identify, stage, and monitor dry eye disease.

The growing prevalence of DED

The TFOS DEWS III Digest Report provides the most up-to-date, consensus-based epidemiology of DED. The global prevalence varies widely depending on methodology, but clear demographic trends emerge.¹

Reported prevalence varies depending on the diagnostic method, ranging from 2.7 to 30.1% with Women’s Health Study criteria, 7.3 to 31.6% with symptom-based definitions, and 4.7 to 62.9% when both symptoms and signs are required.¹ This variability underscores the symptom-sign mismatch commonly seen in practice.

With stricter criteria, prevalence generally falls within 5.4 to 44.2%, starting as low as 4.7% in children and rising to 62.9% in young adult females.¹ Further, meibomian gland dysfunction (MGD), the leading cause of evaporative DED, affects up to 66.7% of some populations, though only about 2 to 23% present with clinically significant disease.¹

For optometry, these figures underscore both the magnitude of DED in the population and the need for precise, reproducible diagnostic strategies, particularly given the significant number of patients who present with either symptoms or signs alone, but not both.

Challenges in traditional dry eye diagnosis

The TFOS DEWS III Digest Report reinforces a long-standing reality for eyecare professionals: diagnosing DED is complicated by a frequent symptom-sign mismatch.¹ Many patients report significant dryness, burning, or fluctuating vision despite minimal observable clinical changes, while others exhibit marked staining, tear film instability, or meibomian gland loss without any subjective complaints.

Eyecare professionals have traditionally relied on a combination of diagnostic approaches, including:

Patient-reported symptom surveys such as the Ocular Surface Disease Index (OSDI) and Dry Eye Questionnaire (DEQ-5)
Tear breakup time (TBUT) measurements
Corneal and conjunctival staining patterns (ex., fluorescein, lissamine green)
Meibomian gland evaluation via expression or meibography
Tear meniscus height assessment with slit lamp or anterior segment optical coherence tomography (OCT)
Tear osmolarity testing

However, as TFOS DEWS III highlights, these assessments can be influenced by examiner technique, inter-observer variability, and environmental factors in the clinic. For example, TBUT values can change depending on fluorescein concentration and instillation volume; staining interpretation can vary between providers; and meibomian gland secretion assessment depends on the amount of pressure applied during expression.

Compounding the challenge, TFOS DEWS III emphasizes that relying solely on symptoms or signs risks misclassifying up to 50% of true DED cases, particularly among younger digital device users, post-surgical patients, and individuals with subclinical MGD.¹ This diagnostic uncertainty can delay treatment initiation, obscure progression, and complicate clinical trial enrollment.

Effective management of DED requires standardized, multimodal evaluation, yet reproducibility continues to pose challenges. Artificial intelligence offers promising solutions by introducing objective and consistent metrics.

The emerging role of AI in eyecare

AI is rapidly transforming how optometrists and other eyecare professionals approach the diagnosis and management of DED. Building on its success in retinal disease screening and glaucoma risk stratification, AI is now being applied to the complex, multifactorial challenges of ocular surface disease.

The rationale for AI adoption in DED is clear: traditional diagnosis suffers from subjectivity, variability, and limited reproducibility, while disease presentation is often heterogeneous. This heterogeneity means that traditional diagnosis, based on subjective patient reports and inconsistent clinical tests, can often lead to variability in how the disease is understood and treated.

AI, on the other hand, can help identify patterns in patient data and offer more consistent, objective assessments, helping to manage the complex, diverse presentations of DED more effectively. AI’s strengths in image analysis, pattern recognition, and predictive analytics directly address these diagnostic pain points by introducing objective, quantifiable metrics that can be tracked over time.²

Examples of AI-powered systems identifying DED

AI-powered imaging tools, particularly in meibography, interferometry, and non-invasive tear film analysis, are demonstrating high accuracy in detecting meibomian gland dropout, tear film instability, and corneal staining patterns—often outperforming manual grading.²

Deep learning models, such as convolutional neural networks (CNNs), can interpret ocular surface images with minimal human input, delivering consistent grading and enabling standardized severity classification across clinics.² Emerging research also highlights AI’s ability to integrate clinical measurements with lifestyle and environmental data to improve prediction and personalization.³

Machine learning models predicting DED-related signs, symptoms, and diagnoses have achieved 60 to 99% accuracy by incorporating heavily weighted lifestyle factors, such as near-work hours, alcohol consumption, outdoor exposure, exercise, and contact lens wear patterns, alongside standard ocular surface metrics.³ This approach recognizes dry eye as, in part, a “lifestyle epidemic” and allows AI to identify modifiable risk factors for earlier intervention.

Utilizing predictive modelling for dry eye patients

The integration of predictive modeling adds further value. AI can forecast which patients are likely to progress or respond poorly to standard therapies, enabling proactive care planning.^2,3

For example, algorithms have been used to predict treatment outcomes based on OSDI/SPEED scores, tear film parameters, and meibomian gland morphology, guiding clinicians toward the most effective therapy combinations for a given patient profile.³

In this evolving landscape, AI is not positioned to replace clinician judgment but rather to serve as a decision-support partner by delivering reproducible data, highlighting subtle trends, and enabling personalized care strategies that extend beyond what human observation alone can achieve.

Table 1: Current AI-driven tools for dry eye diagnosis.^2-4

AI Tool	Clinical Application	Reported Accuracy
AI-Enhanced Meibography	Automated grading of meibomian gland dropout, atrophy, and tortuosity	~74% (with lifestyle data)
Interferometry and Lipid Layer Analysis	Quantification of lipid layer thickness and pattern classification	>84% (with LLT and CL wear history)
Non-Invasive TBUT Mapping	Automated, dye-free TBUT analysis with heat maps	~80% (with lifestyle data)
Ocular Surface Staining Quantification	Pixel-by-pixel grading of fluorescein/lissamine staining	>91% (lifestyle-linked severity prediction)
Predictive Analytics	Forecasting disease progression and treatment response	86.5% (contact lens discomfort prediction)

Ocular surface imaging/analysis technologies

AI-enhanced meibography

MGD is one of the most common underlying causes of evaporative DED.¹ Modern AI platforms use deep learning, such as CNNs, to automatically grade gland dropout, tortuosity, and atrophy with minimal operator input.^2-4

Figure 1: An example of AI morphometric parameter analysis;⁴ AI-segmented glands are skeletonized to allow calculation of morphometric features such as length, width, tortuosity, and gland count. These quantitative metrics move beyond subjective grading, offering objective and reproducible measures of meibomian gland health.

^{Figure 1: Meibography training images©Md Asif Khan Setu et al. Image used under CC BY 4.0.}

Figures 2 and 3: AI segmentation performance on the upper and lower eyelids, respectively.⁴ Figure 2 highlights AI-driven deep learning segmentation on upper eyelids; each row shows the original image, expert-annotated reference, AI-generated segmentation, and an overlay. Arrows highlight areas where glands were missed or falsely segmented.

Even with these minor errors, the AI still performed very well, achieving an area under the curve (AUC) of 0.96 (reflecting near-perfect ability to distinguish gland tissue from background) and a Dice coefficient of 84% (a measure of how closely the AI’s segmentation matches human grading).

AI segmentation performance (upper & lower eyelids)

^{Figure 2: Upper MG segmentation©Md Asif Khan Setu et al. Image used under CC BY 4.0.}

AI-driven deep learning segmentation on upper eyelids

^{Figure 3: Lower MG segmentation©Md Asif Khan Setu et al. Image used under CC BY 4.0.}

Figure 4: Example of AI reliability on poor-quality images.⁴ AI segmentation performance is still reliable in cases of inferior image quality: out-of-focus eyelids, eyelash interference, partial eversion, and even a finger artifact. The algorithm still reliably delineates glands, demonstrating resilience in real-world clinical conditions.

^{Figure 4: MG segmentation©Md Asif Khan Setu et al. Image used under CC BY 4.0.}

Takeaways on utilizing AI-enhanced meibography:

Performance: AI models detect gland dropout patterns with high accuracy and consistently match or surpass the reliability of manual grading, providing clinicians with a reproducible and objective tool they can trust.^2,4
Integration with lifestyle data: Adding lifestyle predictors (e.g., near work, alcohol intake) to meibography features improved model prediction of MGD to ~74% accuracy.³
Clinical impact: Allows optometrists and other eyecare professionals to objectively track gland loss progression, improving patient counseling and timing of thermal or intense pulsed light (IPL) therapies.^2-4

Interferometry and lipid layer assessment

Tear lipid layer thickness (LLT) and stability are critical markers in diagnosing evaporative DED. AI-powered interferometry precisely classifies lipid patterns and quantifies LLT, enabling models that integrate contact lens wear history and LLT data to predict tear instability with over 84% accuracy in key subgroups.³

These objective insights empower proactive care for at-risk populations, including post-refractive surgery patients, high-screen-time users, individuals with MGD, contact lens wearers, and those exposed to dry or windy environments.^2,3

Tear film analysis with AI-based tools

Non-invasive TBUT (NITBUT) with AI mapping

High-speed videography paired with AI segmentation provides automated TBUT analysis without fluorescein instillation.² Advantages include eliminating operator bias, producing color-coded instability maps, and standardizing baseline measurements for longitudinal care.²

In high near-work patients, a NITBUT of <9.0 seconds, when combined with lifestyle factors, predicted the presence of DED with ~80% accuracy.³

Tear meniscus height and volume via AI-OCT

AI can analyze anterior segment OCT scans to measure tear meniscus height with micron-level precision.² These measurements aid in distinguishing aqueous-deficient from evaporative disease, with reduced tear volume more indicative of aqueous-deficient DED, while lower tear meniscus height may also be observed in evaporative forms.¹

This information helps guide therapeutic decisions regarding aqueous support versus lipid-layer-targeted therapy.

Ocular surface staining evaluation

AI-driven staining quantification

Fluorescein and lissamine green staining are longstanding tools in DED, but grading is subjective. AI systems can quantify staining pixel-by-pixel, providing reproducible National Eye Institute (NEI) or Oxford scale scores.²

In clinical studies, AI grading improved diagnostic agreement by up to 15% compared to clinician grading alone.² Lifestyle-integrated models have linked reduced outdoor exposure and higher alcohol intake to increased staining severity with >91% model accuracy.³

Predictive analytics and personalized care

Perhaps the most transformative AI function is predictive modeling, which integrates ocular surface metrics with patient demographics, lifestyle, and symptom scores to forecast disease course and treatment response.^2,3 For example, debilitating dryness in contact lens wearers was predicted with 86.5% accuracy by combining OSDI/SPEED scores, comfortable wear time, and lifestyle metrics.³

AI models can recommend individualized treatment plans, such as optimal artificial tear formulation or thermal therapy candidacy, based on a patient’s risk profile, reducing trial-and-error prescribing.^2,3

Pros, cons, and clinical caveats of using AI in DED diagnosis

Pros

Objectivity and reproducibility: AI minimizes examiner-dependent variability in grading meibography, tear film stability, and ocular surface staining. It provides consistent, quantitative data that is critical for tracking progression and treatment outcomes.²
Early detection and risk stratification: Machine learning models can identify subclinical DED before it becomes symptomatic, enabling preventive interventions.² Lifestyle-integrated models can flag modifiable risk factors such as alcohol intake, outdoor time, and near-work exposure with accuracy greater than 80%.³
Workflow efficiency: AI-powered image analysis can run in the background, generating reports in seconds. This supports same-day diagnosis and management during a single patient visit.^2,3
Personalized care: Predictive analytics recommend therapy pathways tailored to each patient’s phenotype and lifestyle, reducing trial-and-error prescribing and improving patient outcomes.^2,3

Cons

Dataset bias and external validity: Many AI models are trained on homogenous datasets, limiting accuracy across diverse ethnicities, age groups, and imaging systems.^2,3 Results may not translate directly from tertiary eye centers to community optometry settings.
Over-reliance on technology: AI is a decision-support tool, not a replacement for clinical judgment. Risk of “automation bias” exists if clinicians accept AI output without cross-checking against patient history, physical exam, and their own experience and educational judgment.²
Integration and cost barriers: High upfront costs for AI-compatible imaging systems and software licensing.² Smaller practices may struggle with EHR integration or staff training needs.²
Data privacy and security: AI platforms handling imaging and personal health data must comply with HIPAA and regional privacy laws.² Cloud-based analysis introduces potential cybersecurity risks.²
Regulatory and validation gaps: Many AI tools for DED remain in early validation phases and lack FDA clearance.^2,4 Standardized protocols for AI evaluation in ocular surface disease are still evolving.

Key takeaways

Dry eye disease is a common, multifactorial, and often underdiagnosed condition, with symptom-sign mismatch complicating traditional evaluation.^1,2
TFOS DEWS III reports global prevalence of DED symptoms from 8.7 to 65.4% and signs from 14.2 to 70.3%, with female patients and older adults being disproportionately affected.¹
AI tools—including meibography, interferometry, NITBUT analysis, and automated staining quantification—offer objective, reproducible measurements that reduce examiner variability.^2,3
Lifestyle-integrated AI models achieve >80% accuracy in predicting DED risk and can identify modifiable risk factors for earlier intervention.³
Predictive analytics help forecast disease progression and guide personalized treatment plans, improving efficiency and patient outcomes.^2,3
Limitations—including dataset bias, cost barriers, regulatory gaps, and the risk of over-reliance on automation—require careful implementation and clinical oversight.
AI is best positioned as a decision-support partner in clinical practice, enhancing but never replacing the expertise and judgment of the physician.

In conclusion

AI is a powerful adjunct for diagnosing and managing dry eye disease in clinical practice, offering speed, objectivity, and personalization.

However, successful adoption requires critical oversight, workflow integration planning, and ongoing validation to ensure real-world performance matches research results.