Insulin dose [44,67]

The next dosage of anticoagulation [37]

Insulin dose [49]

Crisis support during acute exacerbations of COPDa [52]

Referral advice to manage pain [36]

Medication and nutrition-specific information for chronic liver disease [83]

Clinical reminders for patients with chronic kidney disease [66]

Ostomy management [89]

General treatment decisions for patients with diabetes [55,92]

Adjusting the modality of therapist interaction to manage pain [19]

Peritoneal dialysis management [90]

Monitoring medication adherence [57,75]

Detecting inhaler administration [46]

Monitoring medication adherence [42]

Detecting insulin administration [68]

Improving medication adherence and safety [56,70]

Improving medication adherence and safety [51]

Predicting blood glucose levels or hypoglycemia events [30,40,48,50,60,61,72]

Predicting risk of asthma or COPD exacerbation [31-34,58,62]

Predicting adverse events or classification of the extent of heart failure [35]

Predicting blood glucose levels [49]

Predicting risk of adverse events of heart failure [76]

Identifying heart arrhythmias [59]

Predicting blood pressure [18]

Predicting pain level [73]

Cancer-related symptoms [79]

Postoperative management [87]

Oral anticancer agents [53,88]

Cancer-related symptoms [84]

Chemotherapy-related side effects [85]

Diet, physical activity, and other lifestyles for patients with diabetes [41,75]

Diet or physical activity for patients with heart failure [76]

Various health behaviors for patients with multiple chronic conditions [80]

Diet, physical activity, and other lifestyles for patients with diabetes [82,91]

Physical activity for patients with chronic pain [43,47]

Diet for patients with chronic conditions (multiple chronic conditions, IBDc or IBSd, or chronic kidney disease) [54,64,90]

Various health behaviors for patients with chronic kidney disease [66]

Diet, physical activity, and other lifestyles for patients with diabetes [17,74,92]

Diet or physical activity for patients with cardiovascular diseases [18,77]

Nutrition for patients with cancer [84,86]

Various health behaviors for patients with chronic pain [63,65]

Symptom self-management ability [69]

Treatment adherence and adherence risk [35]

Prediction of ambulation status and independence [39]

Monitoring rehabilitation [38,45]

Monitoring physical activity [18,84]

Monitoring diet [71]

Encouraging expressing emotions through facial and body animations [75]

Emotional support during periods of low moods [52]

Responding to patients’ distress [78]

Identifying psychosocial concerns and providing recommendations [81]

Building emotional attachments [53,83]

Emotional support by recognizing feelings from physiological data (voice, heart rate) [18]

Enhancing psychological flexibility [73,84]

Motivating to support performing self-management [17,74]

Encouraging patients to perform self-management [57]

Motivating to support performing self-management [52]

Motivating and reinforcing the desired self-management activities [63,65]

The SVM classifier implemented in a smartphone app was trained using a fruit database consisting of 10 types of fruits. Each fruit type included 60 images.

SVM classifies food types and volumes and calculates the amount of carbohydrates to help patients control their diet.

Overall accuracy of SVM classification of fruits was 90%.

The average error rate was 6.86%.

The linear regression method was selected to gain a prediction model, and 2 algorithms were tested: gradient descent and normal equation.

The key technique is that the system uses ML to extract the prediction model from the training sample based on user input.

ML algorithm predicts postprandial glucose level by analyzing the diet of patients.

Prediction accuracy using gradient descent and the normal equation was 63% and 73%, respectively.

If a set of blood glucose values is available for a given week, it can be predicted whether the patient will have a hypoglycemic episode in the following week.

ML algorithms predict a hypoglycemia event in the next 24 h using self-monitored blood glucose and medication information.

Prediction accuracy was over 90% in models using RF or SVM.

RF and SVM models had a 91.7% sensitivity and 69.5% specificity.

After incorporating medication information, the sensitivity and specificity were over 90%.

The DL model predicts daily glucose levels based on patient health data, including glucose levels from the day before, diet, physical activity, and weight.

Neural networks used multiple layers of computational nodes to model how mobile health data progressed from one day to another from noisy data.

Prediction accuracy was within 10% of the actual values based on the Clark Error Grid.

An intelligent virtual assistant–based system (VASelfCare) supports medication adherence and lifestyle change, including healthy diets and physical activity.

Patients interact with the virtual assistant that can speak and express emotions through facial and body animations.

Participants (n=20) reported 73.75 of the system usability score on average, which is a borderline rating of excellent.

A conversation AI (My Diabetes Coach) provides personalized support, including blood glucose monitoring, diet, physical activity, medication, and foot care.

Algorithms were tailored according to the clinical targets and recommendations provided by each participant’s health care providers.

Participants in the intervention group (n=93) and control group (n=94) reduced HbA1cg (intervention: 0.33% and control: 0.20%) compared to baseline.

Health-related quality of life utility score was improved in the intervention group (P=.04).

A conversational AI (chatbot) communicates with patients regarding diet, physical activity, and blood glucose and provides personalized feedback based on previous data.

Patients can receive motivational messaging, reducing the difficulty in performing specific tasks and providing triggers needed to act.

AI-powered decision support system (Wellthy CARE) enabled across the platform.

Participants reported reduced HbA1c by 0.49% (n=102), fasting blood glucose by 11 mg/dL (n=51), postprandial blood glucose by 21 mg/dL (n=51), BMI by 0.47 kg/m2 (n=59), and weight by 1.32 kg (n=59) after 4 mo.

ML based on attributable components analysis identifies patterns and relationships between meals and changes in blood glucose levels.

The system (GlucoGoalie) translates ML output into actionable support by generating natural language recommendations for personalized nutritional support to improve blood glucose levels.

In the goal comprehension task, participants accurately selected between 2 nutrition labels 89% of the time.

When choosing between 2 meal images, their accuracy was 49%.

The DL model detects early adherence to once-daily basal insulin based on continuous glucose monitoring and injection data.

Six different detection models were compared according to whether they fused expert-dependent and automatically learned features.

The 3 models based on expert-engineered features reported mean accuracies of 78.6%, 78.2%, and 78.3%.

The model based on learned features reported a mean accuracy of 79.7%.

The 2 models fusing expert-engineered and learned features reported mean accuracies of 79.7% and 79.8%.

Patients can receive personalized recommendations reflecting their health-related social needs for self-management detected by NLP.

The overall engagement ratio with personalization was an average of 0.31, while the engagement ratio without personalization was 0.26.

An AI-based mobile app (BlueStart) provides personalized feedback and coaching to help patients track their medication, sleep, exercise, and other health behaviors.

Patients can view their glucose levels and identify patterns through a continuous glucose monitoring system (Dexcom G6) that synchronizes with the app.

Participants with baseline mean glucose >180 mg/dL (18/52) demonstrated significant improvements in glycemic control after 3 mo.

An AI-based mobile dietary management platform (Auto-Chek Care) collects data from multiple devices linked via Bluetooth and performs integrated analysis.

A DL-based food recognition system (FoodLens) incorporates diet and nutritional data from photographs taken by patients into the platform.

The decreases in HbA1c from baseline to 6-mo in intervention group 1 (–0.32, SD 0.58%) and intervention group 2 (–0.49, SD 0.57%) were significantly larger than those in the control group (–0.06, SD 0.61%)

Intervention groups demonstrated greater weight loss than the control group after 6 mo.

The algorithms are used to perform next-day hypoglycemia prediction in daily life based on the data input from a mobile app (forDiabetes) and portable devices.

RF showed the best accuracy (0.814) and F1-score (0.812) with sensitivity (0.815) and specificity (0.824).

The accuracy of other models ranges from 0.65 to 0.80.

ML algorithms predict nocturnal hypoglycemia based on continuous glucose monitoring data.

The light gradient boosting machine model had the highest predictive performance: accuracy (0.801), specificity (0.802), and F1-score (0.255).

Jump neural network means a feed-forward neural network whose inputs are connected not only to the first hidden layer but also to the output layer.

Neural network predicts continuous blood glucose based on patients’ input of the carbohydrate amount.

Predictions are accurate and close to target (root mean squared error=17.6 mg/dL).

The DSSk, named GlucoP, was designed to help patients in real-time while performing therapeutic corrective actions, including administration of insulin bolus to correct hyperglycemia or intake of carbohydrates in case of hypoglycemia.

The glucose predictor is based on an artificial neural network trained with continuous glucose monitoring profiles.

After perceiving the glucose prediction, 20% of participants decided to revise their initial decision.

Participants (n=12) reported positive opinions about usability (>7 on average out of 9) and described using the DSS as a pleasant experience.

The ML-based ABBAl allows daily adjustment of the insulin infusion profile to compensate for fluctuations in the patient’s glucose level.

ABBA provides personalized suggestions for the daily basal rate and prandial insulin doses based on the patients’ glucose level on the previous day.

ABBA significantly decreased the percentage of time in hypoglycemia and severe hypoglycemia ranges (P<.05), while the percentages of hyperglycemia were increased.

The DRLm adviser can calculate the gain of the meal insulin bolus to help users control the insulin pump or pen.

The DRL adviser provides a personalized insulin bolus adviser to optimize insulin at mealtime.

DRL insulin bolus adviser improved percentage time in target scope (70-180 mg/dL) from 74.1% to 80.9% (for adults) and 54.9%-61.6% (for adolescents) while reducing hypoglycemia.

The DL algorithm predicts at bedtime the probability and timing of nocturnal hypoglycemia based on glucose metrics and physical activity patterns.

Predictions are used to prescribe bedtime carbohydrates in a timely manner.

An area under the receiver operating curve of the model ranged from 0.71 to 0.80 to predict nocturnal hypoglycemic events.

MobiGuide is an AI-based mobile system based on computer-interpretable guidelines for providing personalized decision support to patients having a DSS at the back end and a body area network on the front end.

AI using DSS enables personalized decision support or feedback based on patient-reported blood glucose, ketonuria, diet, blood pressure, and physical activity without medical supervision.

Participants (n=20) reported a higher degree of compliance and satisfaction.

Systolic and diastolic blood pressure were significantly lower in the intervention group (P<.001), compared to the historical cohort group.

ML algorithms predict before an asthma exacerbation occurs based on data, including respiratory symptoms, sleep disturbances, limitation of physical activity, medication use, and measured peak expiratory flow.

3 models using naïve Bayesian, adaptive Bayesian network, and SVM predicted asthma exacerbation occurring on day 8, with the sensitivity of 0.80, 1.00, and 0.84; specificity of 0.77, 1.00, and 0.80; and accuracy of 0.77, 1.00, and 0.80, respectively.

ML algorithms–based system (myAirCoach) conducts short-term prediction of asthma control levels for real-time personalized guidance and long-term prediction of exacerbation risk.

Prediction accuracy of 3 models ranged from 0.79 to <0.86 according to tested cases.

ML algorithms–based system (myAirCoach) conducts an estimation of the risk of asthma exacerbation 7 d ahead based on data.

Prediction accuracy of 4 models was good for predicting exacerbation 7 d in advance.

RF algorithm had an overall greater accuracy, and SVM was effective for predicting the true positive cases.

A conversational AI (Avachat) provides crisis support during acute exacerbation periods and information at the time of diagnosis.

Patients can be motivated to perform self-management and receive emotional support during periods of low moods.

The median system usability score was 73.75 out of 100 (n=8).

ML algorithms–based system (myAirCoach) conducts short-term prediction of asthma control levels and long-term prediction of exacerbation risks in the myAirCoach decision support.

Prediction accuracy of the RF algorithm was the best at 0.80.

SVM and RF classifiers were superior in all cases compared to the AdaBoost and naïve Bayesian.

The DL model, using recurrent neural networks with long short-term memory units and spectrogram features, was tested to monitor medication adherence.

The DL model-based audio signal segmentation approach monitors medication adherence by detecting pressurized metered-dose-inhaler audio events.

Prediction accuracy of the DL model with intrasubject and intersubject settings ranged from 0.92 to 0.94.

ML algorithms classify whether a patient is stable or unstable and allow early warning of asthma aggravation.

Prediction accuracy of both logistic regression and naïve Bayesian algorithm was the best at 0.87.

Data regarding symptoms and self-management abilities were entered into an artificial neural network using a 3-layer model (input, hidden, and output layers).

To ensure stability and reproducibility, a standard feed-forward, multilayer perceptron neural network was used.

The model classified participants into their respective self-management ability tertiles (low, medium, or high).

Prediction accuracy of artificial neural network to predict self-management ability was 0.94 with 21 d of consecutive data.

An interactive cloud-based digital app (myCOPD) predicts exacerbations before 1-8 d based on patient-reported data.

An AdaBoost model showed 35% sensitivity and 89% specificity.

The EasyEnsemble classifier showed 67% sensitivity and 65% specificity.

ML algorithms–based system (HEARTEN Knowledge Management System) classifies patients’ NYHAp class, predicts adverse events, and estimates treatment adherence and the adherence risk regarding medication and overall.

Prediction accuracy of the knowledge management system ranged from 0.78 to 0.95 depending on the algorithms and modules.

Conversational AI–based on proprietary algorithms provides support and tailored coaching to improve self-management and healthy behaviors.

After 6 mo, self-confidence in controlling blood pressure in the intervention group (n=144) was greater than in the control group (n=153).

There was no difference in the effect of reducing blood pressure between the groups.

AI collects data regarding treatment compliance and symptoms and generates 3 types of flags based on the risk.

Patients are encouraged to maintain their current medications, diet, and activities or advised to follow specific suggestions depending on flags.

Older age, a lower number of medications, and being non-Black are associated with higher use of this technology.

ML algorithms were used to predict the presence or not of cardiovascular risk by processing medical records.

The DL algorithm using deep neural networks detects heart arrhythmias by mapping a sequence of electrocardiogram waves to a sequence of rhythm classes.

F-measures were high in the 2 models, 0.83 in the logistic regression classifier and 0.81 in the RF classifier.

The F-measure of deep neural network was 0.83.

ML algorithms–based system (HeartMan) estimates continuous blood pressure, monitors physical activity using the acceleration data, or recognizes motivated, anxious, and depressed feelings from voice and heart rate.

DSS provides exercise plans based on data for patients’ recommendations tailored to patients’ psychological profiles.

The mean absolute error of blood pressure estimation was 9.0/7.0 mm Hg systolic and diastolic blood pressure.

F-measure for physical activity recognition was 0.71. the prediction accuracy of the psychological profile was 0.89.

Participants’ (n=56) self-care behavior was significantly improved, and rates of depression, anxiety, and perceived sexual problems were reduced.

A chatbot (ViK) generates medication reminders and provides personalized responses by interacting with patients.

Patients can build up emotional attachments with the chatbot, contributing to their improved quality of life.

The average medication compliance of patients using the reminder function significantly improved by more than 20%.

A chatbot provides appropriate responses to patients about unfamiliar symptoms that they experienced.

Among 60 questions provided to participants (n=12), 8 (13%) did not match the appropriate topics.

The average score of satisfaction was 2.7 out of 5.

AI-based on-facilitator system (CancerChatCanada) using NLP technology identifies keywords related to patients’ psychosocial concerns and recommends appropriate resources addressing each concern.

Prediction accuracy was 0.797, recall was 0.891, and the F1-score was 0.880.

A virtual assistant using the Amazon Echo Show with Alexa provides tablet-based supportive care software (Nurse AMIEq).

Nurse AMIE monitors lifestyle behaviors, symptoms, and emotional distress and provides timely recommendations.

Participants (n=42) reported high levels of acceptability, feasibility, and satisfaction.

There were no significant effects on psychosocial distress, pain, sleep disturbance, fatigue, physical function, or quality of life.

A knowledge-based chatbot (ChemoFreeBot) interacts with patients regarding chemotherapy-related self-management and side effects through the WhatsApp app.

Participants in the intervention group (n=50) had significantly fewer, less severe, and less distressing symptoms compared to nurse-led education (n=50) and the control group (n=50).

An AI-based virtual platform (Ina) provides ongoing personalized nutritional and symptom guidance via SMS text based on patient-reported data.

A team of live oncology-credentialed dietitians confirms or modifies the guidance if needed.

94% were satisfied with the platform, and 98% reported that the guidance was helpful.

84% and 47% used the advice to guide diet and recommended recipes, respectively.

82% and 88% reported improved quality of life and symptom management, respectively.

A knowledge-based question-answering chatbot (GastricFAQ) provides real-time answers for patients’ self-management after curative gastrectomy.

The overall mean usability score was 4.28 out of 5 (n=56).

The chatbot’s accuracy and F score were 85.2% and 92%, respectively.

A mobile phone text messaging-based chatbot (PENNY-GI) provides tailored medication reminders and promotes medication adherence and toxicity management.

Less than 10% of medication or symptom-related messages were identified as incorrect recommendations for participants (n=40).

Participants reported that medication reminders are useful but found symptom management tools too simple to be helpful.

The multilayered perceptron artificial neural network was used to learn from historical examples, analyze nonlinear data, and hand imprecise information.

An AI algorithms–based mobile app (Well Health) provides the most appropriate therapeutic exercise program by processing data input from patients’ subjective symptom assessment.

Participants (n=161) reported reduced median pain scores from 6 (IQR 5-8) to 4 (IQR 3-6) after using this app.

ML algorithms provide referral advice based on patients’ data to help patients manage their low back pain timely.

Prediction accuracy of 3 models ranged from 0.53 to 0.72 depending on the algorithms and dataset.

A model using boosted tree was the best for predicting referral advice (κ 0.2-0.4).

The algorithm using reinforcement learning was used to address the task of being continuously adaptive.

The ML algorithm–based system (MyBehaviorCBP) analyzes self-reported physical activity logs and is used to generate personalized physical activity recommendations based on the past behaviors of patients.

Participants (n=10) reported increased walking time for a further 4.9 min/d after receiving recommendations from a 5-wk pilot study.

There was no difference in the effect of reducing chronic back pain according to recommendations.

Case-based reasoning system (selfBACK), a branch of knowledge-driven AI, provides weekly personalized self-management recommendations and motivates patients to perform desired behaviors.

The adjusted mean difference in RMDQr score between groups was 0.79 at 3 mo, favoring the intervention group after 3 mo.

The Wysa app, using an anonymous conversational AI, uses a free text conversational interface to listen and respond to patients’ distress by providing evidence-based recommendations.

Patients who used the Wysa app were reported to have improvements in means for depression and anxiety symptom scores with a medium effect size (Cohen d=0.60-0.61).

An intelligent agent used reinforcement learning to learn to progressively refine decisions based on probabilistic trials of new choices with feedback about the response.

The intelligent agent adjusts the modality of therapist interactions and provides recommendations based on response.

A greater portion of participants in the intervention group (n=168) reported improvement in 6 mo in RMDQ (37% vs 19%) and pain intensity (29% vs 17%) than the control group (n=110).

The cloud-based AI app (PainDrainerTM) provides digital information material daily via a tablet, smartphone, or computer.

The AI engine analyzes patient-reported data regarding pain, sleep, work, physical activity, leisure time, and housework; predicts pain levels; alleviates pain; and increases psychological flexibility.

Participants reported significantly reduced pain interference scores at 6 wk (n=41) and sustained at 12 wk (n=34).

57.5% and 72.5% of participants demonstrated statistically significant MIDs for pain and physical function, respectively.

100% and 81.3% of participants demonstrated an improvement in depression and anxiety scores, respectively.

A knowledge-based AI decision support app (SELFBACK) provides weekly tailored self-management recommendations for physical activity and exercise and motivates the desired behavior.

The AI-based app adjunct to usual care did not significantly improve musculoskeletal health compared to control groups at 3 mo (n=294).

ML algorithm monitors medication adherence by tracking wrist motions recognized as a logical sequence of motions.

Prediction accuracy in adherence detection was 0.78 with only 1 sensor worn on either of the wrists.

ML using activity recognition, developed as possibilistic network classifiers, was used to monitor patients’ daily activities, infer whether they took medications, and provide reminders.

The agent encouraged patients to maintain appropriate behaviors or change inappropriate behaviors by evaluating adherence and sending messages.

Prediction accuracy ranged from 0.73 to 0.84 for taking medication, pills, and getting water.

2 ML-based approaches were selected to test: model predictive control and neural networks using a simple feed-forward network.

ML approaches predict and recommend the next dosage of anticoagulation medication by tracking data, including the INRt value.

The mean squared error between recommended and real dosage in the models ranged from 0.0297 to 0.4711.

Intelligent AI implemented in an augmented reality headset manages information related to medication plans, restrictions, and patient preferences and sensor input data to help patients select the right medication and dispense pills in a pillbox following the prescription.

Participants reported the technology was acceptable and feasible.

Older age was associated with less comfort and more hesitancy in using the technology.

AI algorithm using neural networks detects medication administration and whether the patient has followed the required steps of handing the medication device and generates an alert if needed.

Insulin pen administration events were detected with 87.58% sensitivity and 96.06% specificity.

Inhaler administration events were detected with a 91.08% sensitivity and 99.22% specificity.

The J48 classification algorithm and EM clustering were used to classify walking patterns.

ML algorithms analyze and classify patients’ walking strategies, and the data is translated into feedback to help patients with stroke effectively self-manage rehabilitation.

In a 10-m walk test (repeated 6 times), there were significant differences in interstride variation computed in this study between stroke survivors (n=14) and healthy control groups (n=10).

An AI platform using neural network (AiCure) identifies the patient, medication, and confirmed ingestion and provides reminders and dosing instructions.

The average cumulative adherence was 90.5%.

After a 12-wk study, adherence based on plasma drug concentration levels was 100% in the intervention group (n=15) and 50% in the control group (n=12).

The DL algorithm using a back propagation neural network calculates motion data measured by wearable sensors and is applied in the motion recognition procedures of sensors for the rehabilitation of patients with frozen shoulders.

Prediction accuracy ranged from 0.60 to 0.95 according to types of exercise.

2 algorithms were selected to test: logistic regression and artificial neural networks.

ML algorithms predict longer-term functional outcomes and independence for self-care activities at the time of hospital discharge in patients with spinal cord injuries.

Prediction accuracy was similar in the 2 models, as it was >0.85 for ambulation status and ranged from 0.76 to 0.86 for nonambulation outcomes.

NLP-based system (EZNutriPal) extracts dietary information and monitors nutrition intake for patients with chronic diseases from speech and free text.

Prediction accuracy in calorie intake estimation was 89.7%.

NLP is used in health recommender systems to provide tailored educational materials for patients with chronic diseases.

The system can achieve a macro precision of up to 0.970 and overall mean average precision scores of up to 0.628.

The ML algorithm based on supervised learning (a combination of gradient descent, regularization, and recursive elements) was selected to test.

The ML algorithm predicts trigger foods associated with adverse symptoms and is used to provide personalized elimination diets for patients.

After 9 wk, 67.6% of total participants (n=39) reported total symptomatic resolution after the study.

There was significant improvement in symptoms and quality of life.

The AI chatbot (ESTOMABOT) communicates with patients about ostomy management via web chat interfaces.

The overall usability score of the chatbot was 81.5, showing excellent usability.

The AI chatbot (PD AI Chatbot), combined with social media (LINE Application), provides a patient interface that includes content regarding peritoneal dialysis management, clinical reminders, diet, and resources.

The average satisfaction scores were 4.5 out of 5 (n=297).

Infection rates of exit site and tunnel infection were reduced (P=.049 and .02).

The peritonitis rate decreased from 0.98 to 0.8 per 100 patient months.

An AI-based mobile app (KidneyOnline) provides interpretation of disease conditions, lifestyle guidance, regular check-ups, early warnings, real-time answers, and clinical reminders based on patient-reported data.

Patients take photos of their medical records, test results, and clinical prescriptions and upload them onto the mobile app.

The KidneyOnline app reduced the risk of composite kidney outcome and the mean arterial pressure.

The AI chatbot (Lucy LiverBot) provides targeted health information regarding disease, medication, and nutrition and monitors health behaviors via tablet.

The chatbot acts as a social companion and improves patient engagement and self-management.

Lucy LiverBot was perceived as a reliable source of information.

Participants identified the chatbot as a potential educational tool and device that could act as a social companion to improve emotional well-being.