Artificial Intelligence for Health: Real-world Insights and Implementation Challenges
Year: 2025
Type: article
Abstract: Just 28 years ago, Deep Blue − the first IBM’s supercomputer − defeated Garry Kasparov in the chess match of the century.[1] At that time, we were far from imagining that artificial intelligence (AI) could drive our cars, recognize complex patterns in medical diagnostic images, or write code in sophisticated programming languages like R or Python. Today, AI has evolved to such an extent that the leaders of major software companies have called for a pause in its development so that humanity can assess the risks posed by such advanced AI systems. This breathtaking advancement has undoubtedly compelled all fields to turn their attention toward AI, and of course, medicine is no exception. AI is defined as the capacity of machines to imitate human intelligence and behavior. As a broad discipline, it encompasses numerous subfields and types. Key subfields include machine learning, where systems learn and improve from experience; natural language processing (NLP), which enables machines to identify and interact with a human language; Computer Vision, allowing systems to interpret visual data; and Robotics, focusing on the design and use of machines to automate tasks. AI further encompasses distinct types, including Narrow AI, designed to perform specific tasks such as voice recognition; General AI, which can perform any intellectual task that a human can do; and superintelligent AI, which outperforms humans in most economically valuable work. Each of these subfields and types plays pivotal roles in the vast AI landscape, driving innovation and advancement in numerous sectors.[2] However, after the release of software like ChatGPT (OpenAI, San Francisco, USA), Bard (Google, San Francisco, USA), and DeepSeek (DeepSeek, China), large language models (LLMs)-based chatbots have emerged as powerful AI tools for health care, boasting broad capabilities that are truly awe-inspiring. Historically, managing multisystem disorders such as cancer or autoimmune diseases has been challenging due to their complex and often non-specific presentation. However, AI is sparking a revolution in health care now. For several years, research has been conducted on the usefulness of the various AI subfields in the medical arena. This includes areas such as diagnosing and classifying diseases, image interpretation, and optimizing research and logistical processes, such as data extraction and cleaning.[3] With its unparalleled ability to analyze vast amounts of patient data, AI has the potential to streamline disease identification by analyzing intricate patterns invisible to the human eye,[4] predict treatment outcomes,[5] and interpret and categorize clinical notes efficiently.[6] Thus, this groundbreaking technology could revolutionize healthcare delivery and transform the patient experience [Figure 1].Figure 1: Key domains in the real-world implementation of artificial intelligence in healthcare. This figure summarizes the core challenges and opportunities identified in the editorial, including (1) data fragmentation and interoperability limitations, (2) the role of natural language processing in handling unstructured clinical data, (3) model generalizability and risk of bias across populations, (4) ethical and legal frameworks related to privacy and consent, (5) barriers to clinical implementation, and (6) future directions such as AI-assisted diagnostics and multidisciplinary collaborationAt the heart of this transformation lies big data, amassed through multicentric clinical registries, acting as treasure troves for the NLP algorithms. The concurrent growth of clinical data warehouses and national-level initiatives to standardize observational clinical data and make them available centrally is an important point to emphasize.[7] AI needs data, and clinical data are now at a stage where AI can exploit that data. With millions of patient records at their disposal, AI can decode complex disease pathways and refine diagnostic accuracy for several diseases. Currently, we are able to detect dozens of ophthalmological diseases and conditions in retinal photographs automatically using deep neural networks with a sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) above 0.97.[8] NLP approaches will allow us to extract and summarize information from clinical records which are still unstructured, empowering clinical decision support systems. However, as AI progresses at an astonishing pace, concerns regarding its risks and ethical implications have garnered the attention of industry leaders. The call for a pause in AI development reflects a collective recognition of the need for humanity to assess and address these risks proactively. Protecting patient privacy and data security, addressing biases and ensuring fairness, promoting transparency and explainability, maintaining human oversight, obtaining informed consent, and ensuring equity and accessibility are crucial. By engaging in careful evaluation and risk assessment, we can ensure that AI systems are developed and utilized to maximize their positive impact while minimizing potential harm. AI systems are only as good as the data they are trained on. Biases in data collection or algorithm design can result in a system serving only a limited subset of the population, exacerbating healthcare disparities.[9] For example, in the field of ophthalmology, a deep learning model developed to identify diabetic retinopathy (DR) on color fundus photographs had a notable disparity in diagnostic accuracy (73% vs. 60.5%) between lighter and darker skin individuals.[10] Therefore, to ensure fairness and efficacy, it is crucial to ensure that the data training these systems reflect the diversity of the global population.[10] Ethnicity, age, sex, and socioeconomic status influence disease manifestation and response to treatment. Omitting these factors can lead to skewed algorithms that fail certain groups. Therefore, overcoming this challenge will require rigorous efforts in data collection and algorithm design, embracing the full spectrum of patient demographics [Figure 1]. The impact of AI in health care is not monolithic – it differs depending on whether the consumer is a patient or a provider. For patients, AI tools enhance autonomy, access, and health literacy. Applications range from symptom checkers to virtual assistants and wearable-based monitoring tools. NLP-based interfaces help patients interpret lab results, medication instructions, and chronic disease management plans.[6] Yet, the benefits are not universal. Risks such as misinformation, misdiagnosis from unverified AI apps, and the digital divide pose real threats to safety and equity. Healthcare providers, on the other hand, could benefit significantly from AI-augmented tools that support diagnostics, clinical decision-making, and administrative efficiency.[11] From radiologists using computer vision to detect malignancies, to NLP systems that automate clinical documentation, AI promises to reduce cognitive and clerical burden.[12] However, these gains come with caveats: explainability remains a key concern, as black-box models challenge traditional standards of medical accountability. Overreliance on automated tools also risks deskilling the clinical workforce. On a systemic level, AI promises transformation toward predictive, preventive, and personalized medicine. But unless regulatory, infrastructural, and training frameworks evolve in tandem, the sector risks further fragmentation and inequality. Some even argue that AI is reshaping not only labor dynamics but also ethical values and the fabric of healthcare relationships.[13] Addressing these multidimensional challenges thoughtfully will be essential as we integrate AI deeper into the healthcare ecosystem. One particularly promising but complex application of AI in health care lies in community-based eye screening. With an increasing number of regulatory-approved AI tools in ophthalmology – ranging from DR and age-related macular degeneration (AMD) to glaucoma and anterior segment pathologies – AI is poised to transform early detection and triage systems. Deep learning models have achieved remarkable performance metrics, such as an AUC of 0.98 for referable AMD,[14] and have shown capability in quantifying tear breakup time,[15] detecting preclinical keratoconus,[16] and segmenting optic discs for glaucoma evaluation.[17] Yet, implementing these tools in real-world, resource-limited community settings reveals substantial gaps. Unlike controlled clinical environments, health camps often deal with patients who present with multiple co-existing conditions. Current AI tools are frequently trained and validated for single-disease detection, which limits their utility in comprehensive screenings. A person with cataract and DR might only receive attention for one, leaving the other unaddressed. In a recent community screening study led by our team, technician-led triaging enhanced with AI assistance was tested. Over three days and 393 patients, we observed a >20% increase in screening sensitivity as technician familiarity with the AI tool improved –highlighting the importance of the learning curve in human-AI collaboration. Despite promising results, several pragmatic barriers emerged: variability in image quality due to suboptimal equipment, language barriers affecting informed consent and instructions, short patient interaction time, and logistical issues in follow-up coordination (unpublished data). These insights reaffirm that AI’s success in healthcare will depend not only on algorithmic accuracy but also on system integration. For example, our triage framework used a combination of AI image interpretation and basic clinical signs (visual acuity, duration of diabetes, anterior segment assessment) to guide referrals. Patients were flagged for referral based on retinal hemorrhages, optic disc changes, or lens opacities detected by AI. Such multi-pronged approaches reinforce AI’s value as an assistant–not a replacement–for frontline healthcare workers (unpublished data). Furthermore, implementation must be accompanied by robust ethical, regulatory, and infrastructural frameworks. Key concerns include data privacy, transparency of decision-making processes (the “black box” problem), liability in case of AI errors, and the risk of clinicians becoming overly reliant on AI outputs. There is also a need to educate both healthcare professionals and the public on how to interpret and interact with AI-generated insights safely and meaningfully. Equally important is ensuring that patient pathways postscreening are not left incomplete – diagnosis and treatment must be accessible following AI-flagged referrals [Figure 1]. As part of our PROTON initiative,[18,19] we envision developing an integrated data ecosystem – “the PROTON Data Lake” – that combines clinical phenotypes, imaging biomarkers, and patient-reported outcomes to train and validate future AI models. Such a resource, enriched by real-world annotations, will be instrumental in bridging the gap between AI research and its deployment in public health systems. In conclusion, the powerful synergy between AI and medicine holds the key to a transformative future of enhanced diagnosis and management of complex medical conditions. LLMs such as ChatGPT, Google Barts, or DeepSeek could become helpful tools in the daily lives of healthcare professionals. However, responsible and ethical development and deployment of AI are essential to harness its transformative power for the benefit of humanity. The journey toward this future must be marked by diligence, prioritizing diversity and inclusion, and preserving the sacred doctor–patient relationship. By navigating these challenges thoughtfully, we can embrace AI’s potential while safeguarding against potential risks. With these paramount considerations driving us forward, the fusion of AI and medicine emerges as a boundless frontier, poised to revolutionize health care for the greater good.
Source: Delhi Journal of Ophthalmology
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: diamond
OpenAlex ID: W4410967470