Artificial intelligence (AI) has become a part of everyday healthcare. It has transcended its role as an experimental research tool to become a part of healthcare practice, transforming the way diseases are detected, treated, and managed. AI is now seen as a viable solution to help alleviate the strain on health systems around the world. Meanwhile, the speed of adoption has spurred significant concerns regarding bias, transparency and poor regulation. The use of electronic records, electronic imaging, and increasing computing power have increased the rate of adoption in healthcare. Increased demand for services, as well as a shortage of clinicians, has also placed pressure on the need to find efficient solutions. These developments are promising, but they also raise significant concerns about accuracy, fairness, and accountability. Clinicians, policy makers and patients need to understand both sides. This paper explores AI in healthcare. Its main focus is on applications in medical imaging and diagnosis, large language models and the ethical challenges of bias and accountability.
AI in Medical Imaging and Diagnosis
One of the most evident clinical applications of AI is in medical imaging. Machine learning systems can read radiographs, mammograms and pathology slides much quicker than a human reader, and can identify subtle patterns that could be signs of disease. Haug and Drazen (2023) describe how computer reading of radiographs, CT and MRI scans, skin images and retinal photographs has become accepted medical practice, as such images need rapid interpretation and local expertise is often lacking. The real effect is shown in a large Swedish breast cancer screening trial where AI-supported screening detected 6.1 cancers per 1,000 women, comparable to the 5.1 per 1,000 found by the standard procedure (double reading), at the same false-positive rate (Lång et al., 2023). Lång et al. (2023) also found that AI support reduced the screen-reading workload of the radiologists by 44.3%. The results are important because there is a worldwide shortage of radiologists and screening programs rely on timely and accurate reading. For cancer, early detection can be very beneficial since treatment is usually less invasive and more effective when the disease is caught early. Importantly, the AI in this trial did not replace the radiologists; they still had final responsibility for decision making. Additionally, the trial used one type of mammography device and one AI system, and performance can vary with technical factors, so an algorithm has to be checked and monitored where it is used (Lång et al., 2023). Thus, AI can improve detection and relieve workload in imaging, provided it is deployed as a carefully validated aid to human judgment.
Interactive example
How AI Supports a Clinical Decision
A made-up screening case that follows the paper’s point about medical imaging: the AI flags, the specialist decides. Pick a step to see what happens at it.
- Data in. The image comes in for analysis. A screening image, such as a mammogram, is stored as digital data and passed to the AI system. The system learnt from other images, so its results have to be checked in the hospital where it is used.
- AI flags. The AI flags an area to look at. The software marks a spot where the pattern could be a sign of disease. The mark is a prompt to look closer, not a diagnosis.
- Specialist checks. The specialist checks the flag. A radiologist reads the whole image, not only the marked spot, and compares it with earlier images, the patient’s history and the clinical context. The flag stands only if the image supports it.
- Specialist decides. The specialist decides what happens next. The radiologist, not the software, decides the next step, such as a recall for more tests or routine screening, and is responsible for it. In the trial the paper describes, radiologists had the final say.
What if the AI is wrong?
- False alarm. The AI flags healthy tissue. The ring sits on normal tissue. Taken at face value, it could mean a recall and tests the patient does not need, and worry while they wait. The radiologist checks the image, finds nothing that needs follow-up and sets the flag aside.
- Missed finding. The AI flags nothing. The image has a small change that needs follow-up, but the AI does not mark it. A reader who looked only at the flags would miss it. Because the radiologist reads the whole image, the change can still be found and the patient recalled.
This is why the paper treats AI as an aid to human judgment: a model with high overall accuracy can still be wrong about one patient.
A made-up teaching example by Writers Per Hour, built on this paper. It uses no real patient’s image or data, gives no diagnosis and no accuracy figures, and does not show how a particular AI product, or the trial in the paper, works.
Illustration by Writers Per Hour. The specialist keeps the decision, as in the principle of protecting human autonomy in the World Health Organization’s Ethics and governance of artificial intelligence for health (June 28, 2021), and the example says what the AI is for and how it can fail, as the Transparency for Machine Learning-Enabled Medical Devices: Guiding Principles of the FDA, Health Canada and the MHRA (June 13, 2024) recommend.
Large Language Models in Clinical Knowledge and Patient Communication
Large Language Models (LLMs) are a more recent category of AI that are directly applicable to clinical knowledge and patient communication. The systems are trained on large datasets of text and can respond to health-related questions in fluent natural language; one of them, Flan-PaLM, reached 67.6% accuracy on questions in the style of the United States Medical Licensing Examination (Singhal et al., 2023). The AI tools may also help with patient communication: when licensed healthcare professionals compared answers to patient questions posted on a public social media forum, they preferred the chatbot’s answer to the physician’s in approximately 79% of evaluations (Ayers et al., 2023). The findings indicate that LLMs can be used to generate responses to patients’ questions, thereby streamlining the workload for healthcare providers and improving patient understanding. An additional advantage of these tools might be to aid regions with fewer physicians where reliable health information is scarce. However, a web-based forum is not the same as a clinical setting, where a patient’s history, examination findings and safety considerations are key components of every decision. LLMs show impressive medical knowledge and communication ability, but their greatest value lies in supporting clinicians under supervision.
Bias, Safety, and Accountability
AI has many advantages; however, it also poses ethical and safety concerns. One of the key issues is bias. AI models can perpetuate inequities within the health system. Omiye et al. (2023) assessed four popular chatbots and found they occasionally regurgitated discredited race-based information, such as false statements about kidney function and lung capacity, and provided inconsistent answers to the same questions. These errors could reinforce health inequalities if the models are used without critical evaluation by clinicians or patients. Similarly, Haug and Drazen (2023) point out that a model that achieves good results on data that does not accurately represent a group might not do as well on that group. It can also be difficult to see the errors in a model, since many models cannot provide explanations for their decisions. There are also related issues of safety, since a model with high overall accuracy might be incorrect in any particular instance. Singhal et al. (2023) reported that the model achieved 67.6% accuracy on medical licensing-style questions, meaning that it got it right only about two out of three times, and clinicians judged its longer answers to agree with scientific consensus far less often than answers written by clinicians. By contrast, in the breast screening trial described by Lång et al. (2023), the radiologists had the final say on decisions, which made human oversight an important safeguard. Another question that remains unanswered is who is accountable when harm is caused; it can be difficult to pinpoint responsibility among the developers, the hospital and individual clinicians. Privacy is also key, as AI systems heavily rely on vast amounts of sensitive health information, which needs to be safeguarded from misuse. These concerns highlight the importance of a variety of training data, independent evaluation, awareness of limitations and clear standards. Thus, the advantages of AI can only be achieved responsibly when bias, accountability and privacy are treated as rigorously as the technology.
Conclusion
AI is revolutionizing healthcare, transforming disease detection, medical information usage, and doctor-patient communication. In imaging, AI-supported screening has detected cancers at a rate similar to standard double reading while cutting the workload of specialists, which matters wherever specialists are in short supply. Large language models have shown robust medical knowledge and the ability to generate clear, empathetic responses in language-based applications, which could help to reduce administrative burdens. Meanwhile, the evidence suggests that the tools can perpetuate harmful stereotypes, deliver unequal results for different populations, and create issues of responsibility and privacy. The smartest direction to take is therefore one that focuses on how AI can assist humans, not replace them, in providing care. All systems should be tested at their local level, monitored on an ongoing basis and regulated by an ethical and legal framework. Health organizations should also train staff not to accept AI outputs blindly. With these conditions in place, AI can transform the way health systems provide care, making it more accurate, more efficient, and more equitable for patients everywhere.
