Table 1

Descriptive summary of 18 artificial intelligence tools evaluated in AI-RCTs

AI-RCT/AI toolMedical fieldAimDescriptionSoftwares/packages usedPrimary outcomeOutcome classificationMain finding
El-Soll et al20/naPulmonary diseasesPrediction of optimal CPAP titrationA general regression neural network with tree-layer structure (input layer, hidden layer and output layer) was trained to predict optimal CPAP pressure based on five input variables.Neuroshell 2, Ward Systems, Frederick, MDTime of achieving optimal continuous positive airway pressure titrationTherapeuticAI guided CPAP titration resulted in lower time to optimal CPAP and lower titration failure rate.
Martin et al21/Patient Journey Record system (PaJR)Chronic diseasesEarly detection of adverse trajectories and reduction of readmissionsSummaries of semistructured phone calls about well-being and health-concerns analysed by machine learning-based and rule-based algorithms. By detection of signs of health deterioration, an alarm was triggered. Alarms were reviewed by a clinical case manager who decided subsequent interventions.Not specifiedUnplanned emergency ambulatory care sensitive admissionsTherapeuticAI tool allowed early identification of health concerns and resulted in reduction of emergency ambulatory care sensitive admissions.
Zeevi et al22; Popp et al27/naNutrition/endocrinologyPrediction of postprandial glycaemic responseA machine learning algorithm employing stochastic gradient boosting regression was developed to predict personalised postprandial glycaemic responses to real-life meals. Inputs included blood parameters, dietary habits, anthropometrics, physical activity and gut microbiota.Code adapted from the sklearn 0.15.2 Gradient Boosting Regressor classPostprandial glycaemic responsesTherapeuticAI tool accurately predicted postprandial glycaemic responses. Individualised dietary interventions resulted in lower postprandial glycaemic responses and alterations to gut microbiota.
Piette et al23/naBehaviouralImprovement of chronic low back pain by personalised cognitive behavioural therapyA reinforcement learning algorithm is employed to customise cognitive behavioural therapy in patients with chronic low back pain. The algorithm learns from patient feedback and pedometer step counts to provide personalised therapy recommendations.Not specified24-item Roland Morris Disability QuestionnaireTherapeuticNot applicable (protocol of AI-RCT)
Sadasivam et al24/PERSPeCTBehaviouralSmoking cessationA hybrid recommender system employing content-based and collaborative filtering methods was developed to provide personalised messages supporting smoking cessation. Data sources included message-metadata together with implicit (ie, website view patterns) and explicit (item ratings) user feedbacks. Each participant received AI-selected messages from a message database that matched their readiness to quit status.Not specifiedSmoking cessationTherapeuticAfter 30 days, there was no difference in smoking cessation rates, although those receiving AI-tailored computer messages rated them as being more influential.
Shimabukuro et al25/InSightInfectious diseasesSepsis predictionA machine learning based classifier with gradient tree boosting was developed to generate risk scores predictive of sepsis, severe sepsis or septic shock based on electronic health record data. Depending on the predicted risk, an alarm was triggered. Further evaluation and treatment was according to standard guidelines.MatlabAverage hospital length of stayTherapeuticAI-guided monitoring decreased length of hospital stay and in-hospital mortality.
Fulmeret al26/TessBehaviouralReduction of depression and anxietyAn AI-based chatbot was designed to deliver personalised conversations in the form of integrative mental health support, psychoeducation and reminders. Users could enter both free-text and/or select predefined responses.Not specifiedSelf-report tools (PHQ-9, GAD-7, PANAS) for symptoms of depression and anxietyTherapeuticAI-based intervention resulted in reduction of symptoms of depression and anxiety.
Wang et al28 32/EndoScreenerGastroenterologyAutomatic polyp and adenom detectionA deep CNN based on the SegNet architecture was trained to automatically identify polyps in real time during colonoscopy.Not specifiedAdenoma detection rateDiagnosticAutomatic polyp detection system resulted in a significant increased detection rate of adenomas and polyps.
Wu et al29; Chen et al33/Wisense/
GastroenterologyQuality improvement of endoscopy by automatic identification of blind spotsA deep CNN combined with deep reinforcement learning was designed to automatically detect blind spots during EGD.TensorFlowBlind spot rateFeasibilityAI reduced blind spot rate during esophagogastroduodenoscopy
Gong et al34/Wisense/
GastroenterologyQuality improvement of endoscopy by automatic identification of adenomasA deep CNN combined with deep reinforcement learning was designed to automatically detect adenomas during colonoscopy.TensorFlowAdenoma detection rateDiagnosticAI increased adenoma detection rate during colonoscopy
Oka et al30/AskenNutrition/endocrinologyAutomated nutritional intervention to improve glycaemic control in patients with diabetes mellitusParticipants use a mobile app to select foods from a large database (>100 000) of menus, which are analysed with regards to their energy and nutrition content by an AI-powered photo analysis system. The trial will compare dietary interventions based on AI-supported vs standard nutritional therapy.Not specifiedChange in glycated haemoglobin levelsTherapeuticNot applicable (protocol of AI-RCT)
Lin et al31/CC-CruiserOphtalmologyDiagnosis and risk stratification of childhood cataractsA collaborative cloud platform encompassing automatic analysis of uploaded split-lamp photographs of the ocular anterior segment by an AI engine was established. Output includes diagnosis, risk stratification and treatment recommendations.Not specifiedDiagnostic performance for childhood cataractDiagnosticAI tool was less accurate than senior consultants in diagnosing childhood cataracts, but was less time-consuming.
Wijnberge et al35 36; Schneck et al37; Maheshwari et al38 39/EV1000 HPI monitoring deviceSurgery/anaesthesiaPrediction of intraoperative hypotensionA machine learning algorithm to predict hypotensive episodes from arterial pressure waveforms was designed. The model output was implemented as an early warning system based on the estimated ‘hypotension prediction index’(0–100, with higher numbers reflecting higher likelihood of incipient hypotension) and included information about the underlying cause for the predicted hypotension (vasoplegia, hypovolaemia, low contractility).MatlabTime-weighted average of hypotension during surgery/frequency and absolute and relative duration of intraoperative hypotensionTherapeuticThe AI-based early warning system performed different under different clinical settings (ie, elective non-cardiac surgery, primary total hip arthroplasty, moderate to high risk non-cardiac surgery patients).
Auloge et al40/naOrthopaedicsFacilitation of percutaneous vertebroplasty by augmented reality/ artificial intelligence-based navigationA navigation system integrating four video cameras within the flat-panel detector of a standard C-arm fluoroscopy machine was developed, including an AI software that automatically recognised osseous landmarks, identified each vertebral level and displayed 2D/3D planning images on the user interface. After manual selection of the target vertebra, the software suggests an optimal trans-pedicular approach. Once trajectory is validated, the C-arm automatically rotates and the virtual trajectory is superimposed over the real-world camera input with overlaid, motion-compensated needle trajectories.Not specifiedTechnical feasibility of trocar placement using augmented reality/artificial intelligence guidanceFeasibilityAI-guided percutaneous vertebroplasty was feasible and resulted in lower radiation exposure compared with standard fluoroscopic guidance.
Wong et al41/Everion/
Infectious diseasesEarly detection of COVID-19 in quarantine subjectsData from a wearable biosensor worn on the upper arm are automatically transferred in real time through a smartphone app to a cloud storage platform and subsequently analysed by the AI software. The results (including risk prediction of critical events) are displayed on a web-based dashboard for clinical review.Not specifiedTime to diagnosis of coronavirus disease 19DiagnosticNot applicable (protocol of AI-RCT)
Aguilera et al42/naBehaviouralIncrease physical activity in patients with diabetes and depression by tailored messages via AI mobile health applicationParticipants receive daily messages from a messaging bank, with message category, timing and frequency being selected by a reinforcement learning algorithm. The algorithm employs Thompson Sampling to continuously learn from contextual features like previous physical activity, demographic and clinical characteristics.Not specifiedImprovement in physical activity defined by daily step countsTherapeuticNot applicable (protocol of AI-RCT)
Hill et al43/naCardiologyAtrial fibrillation detectionAn atrial fibrillation risk prediction algorithm was developed using machine learning techniques on retrospective data from nearly 3 000 000 adult patients without history of atrial fibrillation. The output is provided as a risk score for the likelihood of atrial fibrillation.RPrevalence of diagnosed atrial fibrillationDiagnosticNot applicable (protocol of AI-RCT)
Yao et al44 45/naCardiologyECG AI-guided screening for low left ventricular ejection fractionA CNN model has been trained to predict low LVEF from 10 s 12-lead ECGs strips from nearly 98’000 patients with paired ECG-TTE data. The final model consisted of 6 convolutional layers, each followed by a nonlinear ‘Relu’ activation function, a batch-normalisation layer and a max-pooling layer. The binary output will be incorporated into the electronic health record and triggered a recommendation for TTE in case of a positive screening result (predicted LVEF ≤35%).Keras, TensorFlow, PythonNewly discovered left ventricular ejection fraction <50%DiagnosticAn AI algorithm applied on existing ECGs enabled the early diagnosis of low left ventricular ejection fraction in patients managed in primary care practices.
  • AI-RCT, artificial intelligence randomised controlled trial; CNN, convolutional neural network; CPAP, continuous positive airway pressure; GAD-7, General Anxiety Disorder-7; LVEF, left ventricular ejection fraction; na, not available; PANAS, Positive and Negative Affect Schedule; PHQ-9, Patient Health Questionnaire-9; TTE, transthoracic echocardiography.