The exciting promise of artificial intelligence (AI) in healthcare has been widely reported, with potential applications across many different domains of medicine [
1,
2]. This promise has been welcomed as healthcare systems globally struggle to deliver the ‘quadruple aim’, namely improving experience of care, improving the health of populations, reducing per capita costs of healthcare [
3], and improving the work life of healthcare providers [
4].
Nevertheless, the potential of AI in healthcare has not been realised to date, with limited existing reports of the clinical and cost benefits that have arisen from real-world use of AI algorithms in clinical practice. This article explores the main challenges and limitations of AI in healthcare, and considers the steps required to translate these potentially transformative technologies from research to clinical practice.
The potential of artificial intelligence in healthcare
A rapidly accelerating number of academic research studies have demonstrated the various applications of AI in healthcare, including algorithms for interpreting chest radiographs [
5‐
9], detecting cancer in mammograms [
10,
11], analysing computer tomography scans [
12‐
15], identifying brain tumours on magnetic resonance images [
16], and predicting development of Alzheimer’s disease from positron emission tomography [
17]. Applications have also been shown in pathology [
18], identifying cancerous skin lesions [
19‐
22], interpreting retinal imaging [
23,
24], detecting arrhythmias [
25,
26], and even identifying hyperkalaemia from electrocardiograms [
27]. Furthermore, AI has aided in polyp detection from colonoscopy [
28], improving genomics interpretation [
29], identifying genetic conditions from facial appearance [
30], and assessing embryo quality to maximise the success of in vitro fertilisation [
31].
Analysis of the immense volume of data collected from electronic health records (EHRs) offers promise in extracting clinically relevant information and making diagnostic evaluations [
32] as well as in providing real-time risk scores for transfer to intensive care [
33], predicting in-hospital mortality, readmission risk, prolonged length of stay and discharge diagnoses [
34], predicting future deterioration, including acute kidney injury [
35], improving decision-making strategies, including weaning of mechanical ventilation [
36] and management of sepsis [
37], and learning treatment policies from observational data [
38]. Proof-of-concept studies have aimed to improve the clinical workflow, including automatic extraction of semantic information from transcripts [
39], recognising speech in doctor–patient conversations [
40], predicting risk of failure to attend hospital appointments [
41], and even summarising doctor–patient consultations [
42].
Given this impressive array of studies, it is perhaps surprising that real world deployments of machine learning algorithms in clinical practice are rare. Despite this, we believe that AI will have a positive impact on many aspects of medicine. AI systems have the potential to reduce unwarranted variation in clinical practice, improve efficiency and prevent avoidable medical errors that will affect almost every patient during their lifetime [
43]. By providing novel tools to support patients and augment healthcare staff, AI could enable better care delivered closer to the patient in the community. AI tools could assist patients in playing a greater role in managing their own health, primary care physicians by allowing them to confidently manage a greater range of complex disease, and specialists by offering superhuman diagnostic performance and disease management. Finally, through the detection of novel signals of disease that clinicians are unable to perceive, AI can extract novel insights from existing data. Examples include the identification of novel predictive features for breast cancer prognosis using stromal cells (rather than the cancer cells themselves) [
44], predicting cardiovascular risk factors and sex from a fundus photograph [
45], inferring blood flow in coronary arteries from cardiac computed tomography [
46], detecting individuals with atrial fibrillation from ECG acquired during normal sinus rhythm [
26], and using retinal imaging to assist an earlier diagnosis of dementia [
47].