INTRODUCTION

The quadruple aims of healthcare are widely recognised as better patient experience, better outcomes, lower costs and team well-being. Recently, a fifth aim of healthcare has evolved, that of improving health equity, to emphasise the importance of social factors as determinants of health.[1] This is in line with HealthierSG, Singapore’s effort to pivot towards population health. Artificial intelligence (AI) has the potential to address the quintuple aims of healthcare by scaling the provision of care through democratisation of professional expertise.

The opportunities for application of AI in healthcare are broad and can span administrative, operational and clinical domains. This article summarises key highlights of the presentations from the recent inaugural International Conference on AI in Medicine, pertaining to the latest developments of AI and future opportunities of using AI in clinical practice.

CURRENT CLINICAL APPLICATIONS OF ARTIFICIAL INTELLIGENCE

The clinical applications of AI include diagnostics, monitoring, treatment, outcome prediction and education. Given today’s developmental context, we highlight cases that illustrate where key opportunities lie.

It is expected that AI will disrupt medical imaging (computer vision) applications such as fundal photography, endoscopy, digital pathology and diagnostic radiology. In Singapore, the most successful example is Selena+, a fundal photography AI solution that democratises ophthalmology expertise to optometrists at the primary care level, which gained overseas adoption and was further validated using hand-held fundal photography units.[2]

Of the 521 US Food and Drug Administration approved Software as a Medical Device AI solutions to date, more than half (393 in total) are radiology applications. Picture archiving and communication systems that collect radiology data are stored in a standardised DICOM (Digital Imaging and Communications in Medicine) format and have provided a fertile ground for computer vision. Imaging examinations that are high in volume, such as chest radiographs and mammograms, are best primed for AI model development and have made way for a plethora of AI solutions.

The incremental value from radiology AI solutions for radiologists is probably marginal when compared to its value to non-radiologists, such as clinicians working in primary care, emergency departments or intensive care units. In these settings, democratising the diagnostic capability of radiologists to frontline clinicians will improve efficiency and quality of care. Furthermore, autonomous application of AI without human-in-the-loop is still not locally acceptable, which means that radiologist manpower demands remain effectively unchanged. Screening mammography would be an exception, since replacing one of two readers could lead to remarkable manpower cost savings.[3]

Computer vision has also been successfully developed for gastrointestinal endoscopy, for which commercially available solutions have been deployed. Artificial intelligence has been used successfully for colon polyp detection and characterisation during colonoscopy.[4] Multiple randomised controlled studies have demonstrated the superiority of AI-assisted colonoscopy over conventional technique for polyp detection[5] in detections of polyps of all nature. Importantly, AI-assisted colonoscopy is cost-effective in routine clinical practice.[6] In addition, anatomical landmark identification, monitoring of speed of scope withdrawal and objective assessment of bowel preparation ensure adequate visualisation for quality assurance.[7] Consideration should be made to patient and clinician autonomy — the right to decide if AI should be used in one’s diagnostic procedure.[8]

EVOLVING APPLICATIONS OF ARTIFICIAL INTELLIGENCE IN CLINICAL CARE

Besides clinical applications, AI can be used for accelerated protein structure-based drug discovery, especially for rare infectious diseases.[9] At the bedside, the ability to synthesise and harness insights from enterprise medical record databases creates new opportunities for clinical decision support across clinical disciplines. These would pave the way for multimodal AI that can effectively integrate clinical information for precision therapy.

For outcome prediction, opportunities in right-siting for care abound. A common example is determining patient disposition at the emergency department. Triaging algorithms have been developed and deployed in large academic centres, for example, PatientFlowNet in Stanford and TriageGo in Johns Hopkins. During the coronavirus disease 2019 (COVID-19) pandemic, AI enabled safe triaging of patients.[10] However, deployment of such models may be limited by non-generalisability, as disease pattern and health policies vary according to geographical location. Local validation is, therefore, required to mitigate such concerns. A recent external validation cohort study of more than 27,000 patients by internists from the University of Michigan highlighted this fact — an enterprise sepsis prediction software had poor discrimination and calibration in predicting the onset of sepsis.[11]

For therapy, dynamic modulation of drug intervention allows personalised dosing for optimising treatment response. Such models rely on ‘small data’ based on dynamic datasets from single patients, rather than large population datasets. For example, the CURATE.AI platform has successfully reduced medication dose to improve response to therapy in oncology and shows promise for chronic disease management.[12]

Recent advances in generative AI (GAI), alongside large language models (LLMs) such as ChatGPT, have generated substantial clinical interest. These LLMs could accelerate the retrieval of key information from vast amounts of unstructured text data in electronic medical records, increasing clinician productivity. With ambient AI, speech conversion of consultation dialogue to clinical notes facilitates automated information organisation. This alleviates administrative demands on clinicians, allowing them to spend quality time with patients. Moreover, structured data entry reduces error, while enhancing efforts in quality assurance, generating new care models and supporting research and innovation. While there are concerns that GAI and LLMs would diminish skills and knowledge assimilation by learners, these could be circumvented with a revamp of the traditional model of medical education. For example, the process of annotating datasets is an important step towards establishing reliable ground truth for model development; opportunities for learning and assessment are created when trainees annotate datasets. Additionally, LLMs will improve the intuitiveness of chatbots, contributing to greater health literacy and nudging patients towards self-activation for health.

MAKING WAY FOR ADOPTION OF ARTIFICIAL INTELLIGENCE

Even as AI technologies are rapidly evolving, the adoption of AI in Singapore healthcare has been somewhat tardy. It is noteworthy that adoption of AI in clinical practice requires stakeholder trust and acceptance. The Universal Theory of Adoption and Use of Technology (UTAUT) model suggests that mainstreaming AI into clinical practice is dependent on model performance (performance expectancy), user competency (effort expectancy), peer utilisation and societal values (social influence), and enabling processes and infrastructure (facilitating conditions).[13] A health-contextualised UTAUT model adapted to the Singapore healthcare system is shown in Table 1.

In ‘performance expectancy’, there is a need to demonstrate clinical efficacy. However, unlike in the biotechnology and medical technology domains, randomised controlled clinical trials of AI interventions are lacking.[14] As AI technology advances, it will become increasingly difficult to evaluate outcomes amidst the concomitant concerns of model drift, bias and hallucinations. In response, AI.Verify, launched by Singapore’s Infocomm Media Development Authority, seeks to address this gap by promoting best practices and standards for AI and by harnessing community-based efforts towards development of AI testing tools to enable responsible AI. Synapxe, an entity under the Singapore Ministry of Health (MOH) and the national healthcare information technology provider, has also developed platforms that facilitate the deployment of AI in radiology (AIM.SG) and GAI through a collaboration with Microsoft.

Long-term acceptance and sustained implementation of AI need healthcare administrators to consider the cost-effectiveness of technology. A recent study on AI clinical decision support tools for dermatology, dentistry and ophthalmology showed that higher model performance did not directly translate to better or cheaper care.[15]

To accelerate the adoption of AI in clinical practice, we have to first determine the efficacy of the models, and several international standards, which recommend best practice that parallels randomised controlled clinical trials, have been put together.[16] Objective assessment will contribute towards clinician and patient trust. To ensure sustained implementation, technical validation, cost-effectiveness and value assessment need to be properly established. To smoothen the deployment of AI in clinical care, simulated environments that examine human–AI interaction for acceptance, and other considerations listed in Table 1, may reduce the risks of poorly tested deployment into clinical settings.

The development of AI models requires access to clean and representative data. Accessible platforms, such as MOH’s TRUST, would enable data sharing across institutions in a secure manner. Better standardisation of data formats would facilitate interoperability. A concerted approach towards data management and the establishment of ground-truth datasets that are common across local healthcare institutions would be ideal. Synthetic datasets can enhance the strength of validation without attendant concerns for patient data confidentiality and safety.[17]

CONCLUSION

In summary, AI has vast potential to change the practice of healthcare. This is most imminent in the areas of diagnosis and outcome prediction, which would aid clinical decision-making. More work needs to be done to educate practitioners and patients. The most pressing challenge that we face today is in guiding our local clinical community to appropriately assess, adapt and adopt AI solutions to optimise patient care. With concerted effort and strong stakeholder buy-in, we can most certainly harness the capability of AI to help us achieve the quintuple aims of healthcare and ultimately, support our national efforts in HealthierSG.

REFERENCES