A user-centered chatbot to identify and interconnect individual, social and environmental risk factors related to overweight and obesity

The Use of Artificial Intelligence–Based Conversational Agents (Chatbots) for Weight Loss: Scoping Review and Practical Recommendations

“…A summary and detailed description of the study characteristics are shown in Table 1 and Multimedia Appendix 3 [ 25 , 26 , 29 - 47 ] and Multimedia Appendix 4 [ 25 , 26 , 29 - 37 , 39 - 47 ]. The chatbot programs included Wakamola [ 30 - 32 ], WaznApp [ 33 ], WeightMentor [ 25 ], SWITCHes [ 34 ], MobileCoach [ 35 ], PathMate2 [ 36 ], and Lark Weight Loss Health Coach AI [ 37 ].…”

“…Most (8/23, 35%) of the studies focused on promoting a healthy diet and exercise, whereas only 4% (1/23) studies focused on a healthy diet, exercise, and stress management ( Figure 2 and Multimedia Appendix 4 ). In all, 11 out of the 14 (79%) planned or trialed experimental studies [ 26 , 29 - 33 , 35 - 37 , 40 , 43 - 45 ] reported program durations that ranged from 1 hour to 12 months [ 30 - 32 ]. Only 1 study mentioned the intention of comparing algorithms to yield accurate behavioral predictions [ 43 ].…”

“…Core functions of the chatbot were to provide personalized weight loss recommendations (20/23, 87%) [ 25 , 26 , 29 - 34 , 37 - 39 , 41 - 46 ] and motivational messages (18/23, 78%) [ 25 , 26 , 29 - 33 , 37 - 39 , 41 - 43 , 45 , 46 ] ( Multimedia Appendix 5 [ 25 , 26 , 29 - 37 , 39 - 47 ]). Only 26% (6/23) studies [ 30 - 32 , 36 , 40 , 43 ] mentioned the use of gamification to enhance user engagement, and 26% (6/23) studies [ 37 , 39 , 40 , 42 , 43 , 45 ] mentioned the use of sentiment analysis to provide emotional support through emotionally appropriate messages ( Multimedia Appendix 5 ). These functions were generally achieved by (1) collecting various user-centric data through chatbot-based self-reports or device-detected metrics, (2) integrating collected parameters using machine learning techniques (including NLP to convert chatbot-collected information for prediction modeling) to predict and generate weight-related recommendations, (3) profiling users according to needs and preferences, and (4) providing personalized chatbot-delivered recommendations.…”

“…The parameters collected included sociodemographic profiles (eg, age, race, ethnicity, education, work status, and income), food consumption, physical activity (eg, intensity, duration, frequency, and type), stress level, sleep (ie, duration), and clinical profiles (eg, presence of specific chronic diseases, medication use, smoking status, heart rate, and blood pressure; Multimedia Appendix 6 [ 25 , 26 , 29 - 37 , 39 - 47 ]). Majority of the parameters were collected through chatbots, except in 3 studies that estimated food consumption using smart refrigerator appliances [ 30 ] and nutritional information provided by retailers (scanning bar codes) [ 38 , 42 ]; 5 that estimated physical exercise type, intensity, and frequency wearable or smartphone sensors [ 33 , 37 , 38 , 42 , 43 ]; 2 that estimated stress levels using plasma cortisol and skin conductance response [ 36 ] and phone detection [ 37 ]; and 3 that measured heart rate and blood pressure [ 36 , 38 , 42 ]. Others that did not mention the use of wearables may have relied on information from in-built sensors of the phone.…”

“…A total of 6 studies reported estimations of chatbot engagement that averaged at approximately 12 minutes a day [ 26 , 45 ], ranging from 4 minutes to 73 minutes per session [ 26 , 30 , 37 , 45 ]. The average daily app use was approximately 71% [ 36 ], with more than 4 conversational turns per day [ 40 ].…”