Artificial Intelligence (AI) in Healthcare
Introduction & Background
A closer look at AI in different medical fields
Artificial intelligence (AI) has entered medicine in many different ways, rather than through one major change. It can appear as a camera attachment in an eye clinic, a second opinion on a mammogram, a language model that shortens a hospital record, or a chatbot that helps a worried parent decide what to do at 2 a.m. This review looks at 26 scientific studies, research reviews, and policy papers covering more than a dozen medical specialties. It asks one main question: using the standards doctors care about, such as diagnostic accuracy, testing in different hospitals, and the ability to repeat results, how much progress has each application really made?
The answer is very different from one medical field to another. This difference is important. In fields that rely heavily on medical images, such as reading photographs of the retina, examining tissue samples, and finding polyps during colonoscopies, several AI systems have shown high accuracy. Some reported area-under-the-curve (AUC) scores above 0.90. These systems were compared with groups of medical specialists rather than just one doctor. More and more have also been tested in different hospitals, using different scanners and patient groups. Even in these areas, progress has not been equal. For years, AI tools for problems in the front part of the eye, such as cataracts, developed more slowly than systems that screen the back of the eye. One reason was the lack of standard ways to capture images. Slit-lamp video recordings eventually gave AI systems a much richer source of information [1]. In other areas, such as predicting side effects from radiotherapy, cancer recurrence, and gynecologic cancer, early results look promising. However, much of the evidence still comes from small studies at single hospitals, with little testing in other hospitals or in real-world clinical settings. In some of the most important areas of medicine, such as regulation, medical training, and hospital leadership, the main questions are not about how well the AI works. They are about the people, systems, and rules needed to use it safely.
The rest of this review looks at each medical specialty separately. It covers eye care, breast and prostate imaging, lung cancer, digestive-system endoscopy, gynecologic and cervical cancer, radiotherapy planning, cancer risk prediction, medical documentation, population health, health apps for consumers, drug discovery, and the rules and training needed to support these technologies. Each section explains what the studies measured, how many patients or images they included, and what the AI systems were compared against. These details help show whether a result is close to being ready for everyday medical use or still needs more testing.
How the research was found and selected
This article is a narrative review of the growing role of AI in healthcare, with special attention to its use in different medical specialties, including cancer care. Its goal was to bring together relevant published research, explain how AI is currently being used in medicine, assess the progress made so far, and describe the problems that could affect its use in everyday healthcare.
Relevant research was found by searching PubMed/MEDLINE, Scopus, Web of Science, and Google Scholar. The search took place over one month, from 1 July 2026 to 1 August 2026. It focused on publications about the use of AI in healthcare and cancer care.
The searches used combinations of terms such as artificial intelligence, machine learning, deep learning, neural networks, generative AI, clinical decision support, medical imaging, cancer diagnosis, radiotherapy planning, side-effect prediction, oncology, and drug discovery. Extra search terms were used to find studies about eye care, breast and prostate cancer, lung cancer, digestive-system endoscopy, gynecologic cancer, and other areas of medical practice.
The review included original research studies, systematic reviews, meta-analyses, and relevant policy papers about the development, testing, or clinical use of AI in healthcare. Special attention was given to studies that reported diagnostic accuracy, patient outcomes, treatment planning, prediction models, and practical problems involved in putting AI technology into use. Publications that were not closely related to the review, repeated reports, or papers that did not provide enough clinical or research details were left out.
Studies were chosen based on how well they matched the aims of the review and how much they helped explain the role of AI in different areas of medicine. Researchers first checked the titles and summaries, then examined the full text of relevant articles. They also looked through the reference lists of important studies and review papers to find other research that might have been missed.
A total of 26 publications were included in the final review. They were chosen to give a broad picture of major developments in clinical AI, including both established uses and newer areas of research. The aim was not to include every paper ever published on the subject. Instead, the review focused on studies that provided useful clinical evidence, showed important technological progress, or discussed major problems with introducing AI into healthcare.
Because this article is a narrative review rather than a systematic review, the research search did not follow a formal systematic-review plan. The authors did not carry out a formal assessment of the risk of bias or a meta-analysis. Choosing which papers to include involved some personal judgment by the authors, so selection bias cannot be ruled out. The findings should therefore be understood as a general review of relevant published research, not as a complete assessment of every available study.
Review
Eye care: cataracts, diabetic eye disease, and glaucoma
Vision loss caused by cataracts, diabetic retinopathy, and glaucoma can often be prevented if these conditions are found early. However, early detection depends on access to an eye specialist, and much of the world does not have enough specialists. This lack of access is where AI has made one of its strongest and most carefully tested contributions to medicine. A photograph of the eye, taken with a hospital-quality retinal camera or a camera attachment connected to a smartphone, can be examined by a trained AI system. This can help find eye conditions such as diabetic retinopathy and glaucoma early and predict their development. AI image analysis can also help screen for cataracts when suitable images of the front of the eye are available.
One study trained an AI model based on EfficientNetV2 using 38,320 usable video frames from 1,812 eyes affected by cataracts. The images were recorded using a slit-lamp device that could be attached to a smartphone. The model graded the severity of nuclear cataracts using the World Health Organization's four-level system. Its AUC scores ranged from 0.923 to 0.967 across the four grades. Its overall accuracy was 0.94, its sensitivity was 0.96, and its specificity was 0.93. These results were close to those of standard slit-lamp examinations [1]. A separate, much larger project led by Google trained an Inception-v3 network using 128,175 retinal photographs. Each photograph was graded three to seven times by a group of 54 eye specialists. The researchers then tested the system using 9,963 images from the United States and 1,748 from France. These images were independently graded by seven to eight certified eye specialists. The AI achieved AUC scores of 0.991 and 0.990 on the two test groups. Its sensitivity reached 97.5% at the screening setting chosen by the researchers. An important finding was that performance stopped improving much once the training set reached about 60,000 images. This suggests that better-quality image grading, rather than simply adding more images, was the main way to improve results [2]. A third study trained an AI system using 39,745 retinal photographs from clinics in China and tested it on a separate group of 8,000 images. It detected glaucoma-related damage to the optic nerve with an AUC of 0.986, a sensitivity of 95.6%, and a specificity of 92.0%. About 8% of the images in this test group were incorrectly marked as positive. A closer examination of errors found that most missed cases involved glaucoma occurring together with severe short-sightedness. This combination also makes diagnosis difficult for human specialists. The researchers also found that 95.4% of the 480 images incorrectly marked as positive showed a real but different abnormality of the optic disc. The most common was a condition called physiologic cupping [3].
All three studies show a similar pattern: AUC scores of 0.92 or higher, with most systems developed and tested within one country, healthcare system, or ethnic group. The main question is no longer whether these systems can read eye images accurately. The evidence suggests that they can. The bigger question is whether they will remain accurate across different ethnic backgrounds, lighting conditions, and camera types in clinics around the world. None of the three studies was designed to fully answer that question.
Breast cancer: from screening images to tissue samples
Breast cancer screening produces huge numbers of medical images that need to be examined by trained specialists. There is also a long-standing problem with missed cancers and false alarms that lead healthy women to have unnecessary biopsies. In 2020, a team from DeepMind and Google Health, working with Imperial College London and several NHS hospital groups, published a widely discussed study in Nature. The researchers used three convolutional neural networks working together. Each had a different job: examining possible problem areas, the breast as a whole, or the complete case. The study brought renewed attention to whether computer software could reduce the workload without making screening less accurate [4].
The system was trained and tested using 121,455 UK screening cases and 22,225 US cases. It achieved an AUC of 0.889 on the UK test group and 0.811 on the US group. At certain settings, it performed better than the average individual radiologist and matched the accuracy of the UK's usual method of having two readers examine each mammogram. However, other scientists strongly criticized the paper because it did not provide complete details about its code and model settings. This led to a disagreement about whether the results could be reproduced. The authors partly defended their decision by referring to patient safety [4]. A wider review of AI across the entire breast cancer care process found similarly strong results in other areas. When 16 studies were combined, AI working alone matched or performed better than individual radiologists in sensitivity, specificity, and AUC when reading digital mammograms. It also reduced the time needed to read breast tomosynthesis images by up to 50%. A randomized trial in Sweden found that having one radiologist work with AI detected cancer at a rate similar to the standard method of using two readers, while greatly reducing the radiologists' workload. One AI system used to read ultrasound images reduced false-positive results by 37.3% and unnecessary biopsy requests by 27.8%, without lowering sensitivity. It performed better than the average results of ten certified breast radiologists. AI systems examining pathology slides also achieved AUC scores of 0.98-0.99 when looking for invasive breast cancer and ductal carcinoma in situ. One model tested in an open competition achieved an AUC of 0.994 when detecting cancer that had spread to lymph nodes, performing better than pathologists working under time limits. Digital scoring tools also showed a strong match with human grading of estrogen receptors, progesterone receptors, and Ki-67. Newer models are beginning to predict BRCA gene mutation status and the risk of cancer returning based on gene-expression patterns, using ordinary H&E-stained tissue slides [5].
Overall, breast cancer is one of the areas of clinical AI that has been tested most thoroughly. At the time of the review, several hundred related devices had already received clearance from the US Food and Drug Administration (FDA). However, most of the strongest results still come from studies using existing records rather than testing AI in real-time clinical practice. The review authors also raise an important fairness concern. Because training data still contains more information about majority populations, AI systems developed from these data could make differences in breast cancer outcomes worse for groups that are underrepresented in research [5].
Prostate cancer and the pathologist's microscope
When examining a prostate biopsy, a pathologist gives the cancer a Gleason score. This score plays an important role in choosing treatment. However, grading is not always straightforward, and even experienced pathologists sometimes disagree about the same case. This makes prostate cancer a useful area for AI research. Researchers have trained systems on large amounts of data and tested whether they can match the opinions of groups of experts rather than just one specialist.
One research team developed a fully automated grading system based on an extended U-Net model. It used a partly automated labeling method that learned from the text in pathology reports instead of relying on the usual detailed, manually marked image areas. This helped researchers use more training data. When tested against the agreed judgment of three pathologists on 535 biopsies, the system achieved a quadratic Cohen's kappa score of 0.918. In a separate study comparing different observers, it performed better than 10 of 15 independent pathologists. Its kappa score was 0.854, compared with a median of 0.819 for the expert group. It also performed reasonably well on an outside Swiss dataset, where the images had been scanned and stained differently. Its kappa scores ranged from 0.707 to 0.723. Most mistakes happened when distinguishing Grade Group 2 from 3 and Grade Group 4 from 5. These are also the distinctions where human specialists have the most difficulty agreeing [6]. A much larger project trained a multiple-instance-learning system using 44,732 complete pathology-slide images from 15,187 patients. The data covered prostate cancer, basal cell carcinoma, and breast cancer in lymph nodes. The total amount of image information was about 88 times greater than that of the entire ImageNet dataset. The system achieved AUC scores of 0.986 for prostate cancer, 0.986 for basal cell carcinoma, and 0.965 for breast cancer in lymph nodes. Adding a recurrent neural network to combine results increased the prostate cancer score to 0.991. When the system was set to detect every positive case, it could safely remove 65-75% of slides from manual review by pathologists. Accuracy fell by about 3 percentage points when images came from a different scanner and by about 6 points when outside consultation slides from other hospitals were used. This drop was smaller than the decline seen in a similarly trained model tested outside its original, carefully selected research dataset. That model's AUC fell from 0.930 to 0.727. This suggests that training on large amounts of less carefully organized, real-world data may help AI work better in different settings than training on small, highly controlled research datasets [7].
These are large studies with careful research methods. Their results suggest that a practical approach may soon be possible: AI can handle routine cases, allowing pathologists to spend more time on difficult ones. The main remaining problem is making sure the systems work consistently in different hospitals. It is also important to note that several authors of the larger study reported financial connections with a commercial company that develops computer-based pathology tools. This conflict of interest should be considered when judging the findings [7].
Lung cancer: finding it in scans and examining it under a microscope
Lung cancer causes more deaths worldwide than any other cancer. Screening with low-dose CT scans has been shown to save lives, but it is difficult to find cancer early without sending too many people for unnecessary follow-up tests. AI has been studied at two stages of lung cancer care: reading the screening scans and, more surprisingly, examining the biology of a tumor directly from a pathology slide.
Predicting the genetic features of a tumor from what it looks like under a microscope, without performing a genetic test, would be a major step beyond simply identifying cancer. However, this research is still at an early stage. Some of the rarer mutations were tested using only a few positive slides, which creates a risk that the model learned details specific to its training data rather than general patterns. The model also could not predict an important mutation type called ALK rearrangement [8]. This work gives us an idea of where pathology AI may go in the future, rather than providing a tool that is ready for everyday use today.
One widely discussed study in Nature Medicine developed a complete three-dimensional deep-learning model. When an earlier scan of the same patient was available for comparison, the model estimated cancer risk from a chest CT scan more accurately than six practicing radiologists. It reduced both false-positive and false-negative results compared with assessment by radiologists alone. Another study trained an Inception-v3 network using 1,634 complete pathology-slide images from The Cancer Genome Atlas. It classified lung tissue as normal, adenocarcinoma, or squamous cell carcinoma. The system achieved an AUC of about 0.97 when telling the two cancer types apart and 0.99 when distinguishing cancer tissue from normal tissue. These results were similar to those of three independent lung specialists who examine tissue samples. The system also performed well in three separate test groups, with AUC scores ranging from 0.83 to 0.98. These groups included frozen tissue samples, paraffin-preserved samples, and biopsies. In a second step, the researchers retrained the system to predict the mutation status of ten commonly mutated genes using only the appearance of the tissue, without any genetic sequencing information. The system could predict six genes, including STK11, EGFR, and TP53, with AUC scores ranging from 0.733 to 0.856. When tested on an outside group of 63 slides involving EGFR, it achieved an AUC of 0.75 [9].
Digestive health: finding problems the human eye can miss
Colonoscopy is very effective at preventing colon cancer, but only when doctors find the precancerous growths called polyps. Human specialists regularly miss some of these growths, especially small or flat ones. This is a task that suits computers well because they do not get tired or lose concentration during a long procedure. The research evidence in this area is particularly strong because it includes randomized clinical trials, not just studies looking back at earlier patient records.
A major review of research on endoscopy reported strong results across several parts of digestive healthcare. AI detected Helicobacter pylori infection with a combined sensitivity of 87% and specificity of 86% across studies. It detected early stomach cancer with 92.2% sensitivity. When results from 21 randomized trials involving 18,232 patients were combined, AI reduced the rate at which precancerous growths were missed by about 55%. Models using cholangioscopy and endoscopic ultrasound reached 90-99% accuracy when identifying cancerous bile duct narrowings [10]. A more focused analysis combined five of these randomized trials, involving 4,354 patients. Four trials took place in China and one in Italy. It found that colonoscopies using AI detected adenomas in 36.6% of patients, compared with 25.2% using standard colonoscopy alone. This was a risk ratio of 1.44. The researchers rated the evidence as moderate quality using the GRADE system. The improvement remained regardless of the polyp's size, location, or shape. This suggests that the software may notice different visual clues from those normally used by human specialists, rather than simply copying what they already do well [11]. Another system designed to detect polyps during live procedures was trained on 8,641 manually labeled images. It was tested using VGG19, ResNet50, and similar computer models. It achieved 96.4% accuracy and an AUC of 0.991. Its ability to mark the exact location of a polyp, measured using the Dice score, was 0.83, compared with 0.55 for an earlier similar method. It processed video at 98 frames per second, about four times faster than needed for genuine real-time use. In a test using real procedure videos, three expert colonoscopists found 45 polyps with the AI system's visual guidance, compared with 36 without it. The AI system also independently identified all 28 polyps that the original colonoscopist had actually removed. The study reported false-positive rates of about 5-7% for individual video frames in its other video tests. However, it did not report the number of false alarms per procedure in the test involving the three colonoscopists [12].
This is one of the strongest examples of clinical AI outside eye care. The evidence includes large randomized trials, shows that AI can improve on existing human performance, and demonstrates that the technology is fast enough to work during a live procedure. The main concern is that most of the research came from a small number of locations. Four of the five randomized trials took place at single hospitals in China, and the control groups had relatively low polyp-detection rates. More research is needed to confirm whether the same benefits will occur in other healthcare systems and in hospitals where doctors already have higher detection rates [11].
Women's health: cervical and gynecologic cancer
Cervical cancer is one of the cancers that can be prevented most effectively. However, prevention depends on regular screening systems that have enough resources to provide cell testing, HPV testing, and colposcopy. Even wealthy countries can struggle to provide these services, and they are often unavailable in other parts of the world. AI has been studied across almost the entire process of gynecologic cancer care, from the first HPV test to diagnosis, cancer staging, and radiotherapy planning.
A large research review covering more than 30 individual studies found that a machine-learning decision system could identify women at high risk of serious precancerous changes, known as CIN2/3 or worse, with 94.32% specificity using extended HPV genotyping. The review also followed the development of computer-based cervical cell screening, from early commercial systems such as PAPNET in 1992 and ThinPrep in 2004 to modern deep-learning systems. Depending on the dataset, newer systems reached accuracy levels ranging from about 93% to above 99%. In one prospective study involving 703,103 women, AI-supported cell screening increased the detection of precancerous cells by 5.8 percentage points compared with manual examination alone [13,14]. An AI-supported system for examining the cervix using colposcopy agreed with the final tissue-test results in 82.2% of cases, compared with 65.9% for standard colposcopy. AI models using pelvic MRI images achieved Dice scores of 0.8-0.98 when marking tumor areas. They also achieved an AUC of up to 0.921 when predicting whether cancer had spread into the tissue around the cervix, and an AUC of 0.933 for the best combined model predicting cancer spread to lymph nodes [14]. A separate analysis combined 12 studies involving 793 patients. It found that special measurements taken from MRI images, known as radiomic features, predicted lymph node cancer spread with a combined sensitivity of 80%, specificity of 76%, and AUC of 0.83. This was better than traditional imaging based mainly on the shape and appearance of tissues, which historically had a sensitivity of only 38-56%. However, the same analysis found that the studies had an average Radiomics Quality Score of just 33.8% of the highest possible score. None had been designed as a prospective study, and only two of the 12 had been tested using data from outside their original research settings [15].
The range of applications being studied is impressive. It includes screening, diagnosis, cancer staging, and even planning the areas that need radiotherapy. Some results, such as the 5.8-percentage-point improvement in screening sensitivity among more than 700,000 women, are supported by research involving very large numbers of patients. However, much of the evidence still comes from studies that look back at existing records or involve only one hospital. The research quality assessments also suggest that the results are promising but not yet conclusive. For now, these tools should help doctors make decisions rather than replace their professional judgment.
Radiotherapy: planning treatment and predicting side effects
Radiotherapy requires a careful balance. Doctors need to use enough radiation to control a tumor without damaging too much of the healthy tissue around it. Predicting which patients will experience serious side effects from treatment, and planning treatment to reduce those risks, has traditionally depended on doctors' experience and general rules about radiation doses and the amount of tissue exposed. These are exactly the kinds of patterns that machine learning may be able to recognize.
A major review examined 53 studies that used traditional machine-learning methods to predict side effects from radiotherapy. These methods included logistic regression, random forests, support vector machines, and artificial neural networks. The studies covered eight cancer sites, including 13 head-and-neck studies, 15 lung studies, and 16 prostate studies, with smaller numbers in other areas. The reported AUC scores varied widely, from about 0.6 to 0.9. The review found a consistent and worrying pattern: almost all models with the highest scores, above 0.85, had been developed using small or medium-sized groups of fewer than 150 patients. This may mean that the models were memorizing details of their training data instead of learning patterns that would work for other patients. Genetic factors known to affect how sensitive people are to radiation were included in only two of the 53 studies, even though research has clearly shown that these factors matter. The review authors recommend that future studies make wider use of the TRIPOD reporting checklist so their methods and results can be compared more easily [16].
The early results suggest that AI may be useful for predicting radiotherapy side effects, but much more work is needed. Before these models can reasonably guide radiation oncologists when choosing treatment doses, researchers need larger studies involving several hospitals. The models must also be tested on patients whose data they have never seen before. This need for testing outside the original research setting is a problem that appears again and again throughout this review.
Cancer risk, recurrence, and survival: earlier machine-learning methods
Before the recent rise of deep learning, an earlier generation of machine-learning methods was already being used to answer important questions about cancer. These methods included artificial neural networks, decision trees, support vector machines, and Bayesian networks. Researchers wanted to know whether computer programs could use a patient's medical, imaging, or genetic information to predict who might develop cancer, whose cancer might return, and who might survive the disease.
A broad review published in 2015 examined many of these earlier studies and organized them into tables based on what they were trying to predict. Some of the reported results were impressive. An artificial neural network used to predict breast cancer risk, based on 48,774 mammogram findings, achieved an AUC of 0.965 when tested using ten-fold cross-validation. A support vector machine using a pattern of activity from 50 genes predicted breast cancer survival with 97% accuracy in a group of 295 patients. Another support vector machine model predicted cancer recurrence with 84.6% accuracy and performed better than both neural networks and Cox proportional-hazards regression. A model using both labeled and unlabeled data was tested on a SEER-based database of 162,500 patients and performed better than the support vector machine and neural network methods. However, the same review openly described an important weakness. One study claimed 100% accuracy when predicting the return of oral cancer, but it had tested the model on only 86 patients. Most of the studies predicting recurrence or survival included fewer than 300 patients in total. They also had little or no testing across different hospitals, and researchers did not always explain clearly how they chose the information used to make predictions [17].
This earlier research is an important starting point for the newer deep-learning imaging studies discussed elsewhere in this review. Its main lesson is still relevant today: studies using small groups of patients from a single source can produce very high accuracy scores, but those results may not hold up when the models are used with different patients. The field has had to learn this lesson again at each stage of AI development.
Reducing paperwork: AI and medical records
Ask almost any doctor what causes work-related stress and burnout, and reviewing medical records is likely to be mentioned. The average doctor spends about 1.5 hours each day working through patient records, and 75% of burnout cases identify electronic health record tasks as a major cause. Large language models, the technology behind modern AI chatbots, are now being tested as a way to shorten and simplify this large amount of paperwork.
Early studies have shown encouraging results. Language-processing tools such as BARTcnn and GPT-3's DaVinci model shortened long brain-imaging reports to less than 20% of their original length. A study published in Nature Medicine found that GPT-4, adapted through in-context learning, performed better than physicians on some measures of summary quality and avoiding invented information. Another approach, called SPeC, uses soft prompting to make language-model responses more consistent from one medical note to another. This matters because inconsistent results are one of the biggest practical problems when deciding whether an AI-generated medical summary can be trusted [18].
There are real benefits, but there are also serious concerns. The average length of an outpatient medical note has increased by 60% since 2009, and more than half of the content is repetitive. This creates problems even for large language models, which can only process a limited amount of information at one time. Research has also repeatedly shown that training data does not represent some population groups fairly. Differences in pulse oximeter readings based on skin tone are one well-known example of this wider problem. There are also unresolved questions about whether patients have properly agreed to the large-scale reuse of their medical information. These concerns are similar to those raised in the 2017 controversy over how DeepMind shared data through the NHS Streams project. Researchers have also shown that some language models can be vulnerable to membership-inference attacks. These attacks can reach up to 80% precision, meaning that someone may be able to manipulate a model into revealing parts of the patient information used to train it. These problems do not mean AI should never be used to help with medical records. They do mean that privacy and fairness need careful and ongoing attention as these systems are introduced [18].
Managing long-term diseases across large populations
Diabetes, high blood pressure, and chronic obstructive pulmonary disease (COPD) account for a very large share of healthcare spending around the world. More research is now looking at how AI could help manage these conditions across entire populations, rather than simply diagnosing one patient at a time. Instead of testing a single computer program, a recent study examined the research field itself. It searched through 758 publications and included 341 in its analysis. These papers covered 775 institutions in 55 countries, from the beginning of the field through August 2024 [19].
The results showed that research was concentrated in a small number of countries. The United States produced 50.7% of all publications and had the highest total citation score. China came next, with 11.1%. The research mainly covered four topics: diagnosis, healthcare delivery, telemedicine, and the technology behind these systems. Diabetes, high blood pressure, and COPD were the most frequently studied conditions. Among the newest papers, the phrase "machine learning" was the most common emerging research keyword [19].
Two findings raise concerns. First, although the number of publications has grown almost exponentially since 2013, the field's total local citation score, which gives an idea of its influence within its own research area, has actually fallen in recent years. This suggests that the amount of research has grown faster than its scientific quality or influence. Second, researchers in Africa and South America contributed very little in terms of both publications and international research partnerships. This lack of global representation is important to remember when someone proposes using an AI tool developed mainly with US and Chinese data in a healthcare system with very different patients and resources [19].
AI in your pocket: symptom checkers and menstrual cycle apps
People were already using AI-based health products long before hospitals began adopting these technologies. Examples include websites and apps that check symptoms and menstrual cycle trackers used by hundreds of millions of people around the world. These products are used outside hospitals and doctors' offices. Research suggests that they have not been tested as carefully as AI systems that examine medical images in clinical settings.
A widely discussed study published in the BMJ tested 23 publicly available symptom checkers using 45 standard patient examples. It carried out 770 diagnosis checks and 532 assessments of what action the patient should take. The correct diagnosis appeared as the first result in only 34% of checks and within the top 20 possible diagnoses in 58%. The recommended next step was appropriate in 57% of cases. Results depended strongly on how urgent the symptoms were. The tools correctly recommended urgent care in 80% of genuinely emergency cases, but their accuracy fell to just 33% when people could safely look after themselves. This means the tools often advised people with mild symptoms to seek formal medical care when it was not necessary [20]. A separate review of menstrual cycle-tracking apps searched 150 articles and included 49 for full review across four research areas. It found that calendar-based predictions of ovulation ranged in accuracy from 17% to 89%, depending on the method used. The FDA-cleared Natural Cycles app also uses basal body temperature readings. Its typical-use Pearl Index was 7.0 contraceptive failures per 100 person-years, compared with 24 for fertility-awareness methods in general. However, in a dataset of 1.4 million eligible menstrual cycles, the researchers could not identify an ovulation date for 665,603 cycles. About three-quarters of these cycles were missing temperature readings for at least half of the cycle. None of the apps assessed using a standard quality checklist scored more than 13 out of 15 possible points. Missing information on this scale matters for health decisions. Natural Cycles has reported a typical-use Pearl Index of 6.9-7.0, compared with about 1.0 with perfect use. The difference is understood to be caused in large part by incomplete everyday record-keeping, rather than by mistakes in the algorithm alone. When temperature readings are missing, the app must use a more cautious backup method that is closer to calendar-based prediction. This method is less reliable at identifying the exact day of ovulation. As a result, the overall typical-use Pearl Index probably combines many different levels of individual risk. People who record their information consistently may get results closer to perfect-use performance, while people who enter little information may face risks closer to those of calendar-based methods in general. None of the reviewed studies reported failure rates separately according to how complete the users' data was. This is an important missing piece of information when assessing how safe the app is for people who do not record their information regularly [21].
It is important to remember that the symptom-checker study was published before today's more advanced large-language-model chatbots became available. Its findings may not reflect what newer tools can do. However, both studies point to the same basic problem: consumer health AI has not been tested as carefully as AI used to examine medical images in hospitals. Both areas of research also have gaps in the groups of people represented in the data. Studies of menstrual cycle apps, in particular, include too many white, European, and highly educated users compared with the wider population.
Creating new medicines from scratch
Developing a new medicine using traditional methods is slow and extremely expensive. It usually takes 10 to 20 years and can cost up to $2.6 billion to bring one new medicine to patients. Generative AI, the broad type of technology used to create images and text, has been tested as a way to shorten this process. Instead of simply searching through existing collections of chemical compounds, these systems try to design completely new possible medicines.
In an important demonstration published in 2019, researchers combined a generative autoencoder, a tensor-train decomposition step, and a reinforcement-learning process. The system was rewarded using a self-organizing map. Its goal was to design new small-molecule drugs that could block DDR1, a kinase involved in fibrosis. Starting with no existing drug candidates, the complete process generated possible chemical structures, produced them in a laboratory, and tested them in cells. It produced six real compounds in just 21 days. Four of the six showed measurable biological activity against the target. Two had IC50 values of 10.3 and 5.8 nanomolar, which were strong enough in cell-based tests to justify further research. The best candidate also showed favorable pharmacokinetic results when tested in mice [22].
This is one of the earliest published examples of an AI-designed drug candidate making it all the way to testing in living biological systems. The speed is impressive compared with the usual drug-development process, which often takes more than a decade. However, this was just one successful demonstration involving one drug target and a small number of possible medicines. The researchers did not directly compare the AI approach with a traditional drug-discovery project working on exactly the same problem. Three authors of the paper also reported having financial interests in the generative AI technology used. This should be considered alongside the otherwise promising results [22].
Rules, training, and leadership: the people and systems needed to support AI
None of the AI tools discussed above will make much difference if regulators cannot approve them safely, doctors are not trained to question their results, or hospitals are unwilling to use them. Several studies outside the clinical-trial literature examine these wider issues. They look at the rules, training, and workplace support that determine whether AI can actually reach patients.
US regulators have proposed a framework called the Predetermined Change Control Plan for AI and machine-learning software that continues to learn after it has been introduced. The framework requires manufacturers to explain in advance exactly how their systems are allowed to change and how those changes will be checked. They do this through a defined SaMD Pre-Specifications (SPS) and Algorithm Change Protocol (ACP). This approach is intended to avoid requiring a completely new regulatory review every time the software is updated [23]. Medical education researchers have separately suggested that future doctors will need three new skills. First, they must act as Evaluators, deciding when it is appropriate to use an AI tool. Second, they must act as Interpreters, understanding where the system's mistakes are likely to come from. Third, they must act as Communicators, explaining AI results to patients in simple language. The researchers referred to Microsoft's Tay chatbot from 2016 as a warning. Within 24 hours of being released to the public, it began producing offensive content, showing what can happen when AI systems operate without proper supervision [24]. Researchers have also reported a pilot project using adaptive e-learning software in medical education. However, evidence supporting these kinds of educational tools is still limited, and most studies have not included strong control groups [25]. Another study examined general leadership and workplace conditions rather than AI or digital health. It surveyed 2,813 doctors at a major US academic medical center, with a response rate of 72.2%. This review applies its findings by comparison to the introduction of AI in hospitals. Researchers found that the quality of a doctor's supervisor, measured using a validated scale ranging from 12 to 60 points, was strongly linked to both burnout and job satisfaction. For every one-point increase in a supervisor's score, the odds of physician burnout fell by 3.3%, while job satisfaction increased by 9.0%. Leadership scores alone statistically explained about 11% of the differences in burnout and 47% of the differences in job satisfaction [26].
Together, these findings highlight something that can easily be overlooked when reading headlines about AI accuracy. Even when an algorithm performs extremely well in a research study, safely bringing it into everyday healthcare depends on regulation, training, and good local leadership. In other words, successfully using AI in medicine is as much a challenge involving people and organizations as it is a technical challenge.
Conclusions
Looking at all the studies together reveals a clear difference in how far various types of medical AI have developed. AI that examines medical images, including eye photographs, mammograms, pathology slides, and colonoscopy videos, is the most advanced area. These systems often perform as well as or slightly better than medical specialists. More of them are also being tested across several hospitals. AI for radiotherapy planning, gynecologic and cervical cancer, and predicting cancer risk is moving forward too, but these areas are less developed. They still depend on smaller patient groups and have not been tested enough in real-world settings or outside their original research hospitals. Consumer health tools, such as symptom checkers and menstrual cycle trackers, have the weakest research evidence. AI used to support hospital work, including summarizing medical records and analyzing population health, may be useful, but it raises unresolved questions about bias, patient consent, and the quality of the information used. Generative AI for designing new medicines is still at the early demonstration stage. Individual experiments have produced exciting results, but there is not yet a reliable record of repeated success.
The main lesson is that, in most areas, the biggest challenge is no longer simply whether the AI algorithms can work. Many already perform at levels similar to trained medical specialists. The harder task is proving that these results remain reliable when the systems are used in different hospitals and with different groups of patients. At the same time, healthcare needs strong rules and better medical training to make sure AI is used responsibly. Hospitals also need good leadership, supportive workplaces, and trust between staff and institutions. These human factors are essential if AI is to become a useful part of everyday medical care. Based on the research reviewed here, the future of AI in medicine looks very promising, but there is still important work to do before it can reach its full potential.
