Adopting Artificial Intelligence in Public Healthcare: The Effect of Social Power and Learning Algorithms

Abstract

Although the use of artificial intelligence (AI) in healthcare is still in its early stages, it is important to understand the factors influencing its adoption. Using a qualitative multi-case study of three hospitals in China, we explored the research of factors affecting AI adoption from a social power perspective with consideration of the learning algorithm abilities of AI systems. Data were collected through semi-structured interviews, participative observations, and document analysis, and analyzed using NVivo 11. We classified six social powers into knowledge-based and non-knowledge-based power structures, revealing a social power pattern related to the learning algorithm ability of AI.

Artificial Intelligence Applications in Health Care Practice: Scoping Review

“…Most (33/45, 73%) of the articles were from North America and Europe [ 46 , 47 , 49 - 55 , 57 , 58 , 61 - 63 , 67 - 70 , 73 - 87 ], of which most (18/33, 55%) were from the United States [ 46 , 47 , 49 - 52 , 54 , 68 , 73 - 77 , 79 - 81 , 84 , 87 ]. The greatest number of AI systems were implemented either in hospital-wide settings (6/45, 13%) [ 50 , 55 , 56 , 65 , 74 , 80 ] or in radiology (6/45, 13%) [ 53 , 56 , 66 , 68 , 73 , 76 ]. Most (27/45, 60%) of the studies were authored by a multidisciplinary team [ 46 , 47 , 50 - 55 , 58 , 59 , 61 , 62 , 64 , 67 , 69 , 70 , 72 , 74 , 75 , 78 - 80 , 82 , 86 - 89 ], with clinical and IT or informatics backgrounds being the most common combination (9/27, 33%) [ 47 , 50 , 55 , 61 , 70 , 74 , 79 , …”

“…The most common AI model was a symbolic or knowledge-based model, reported in nearly half (22/45, 49%) of the reviewed studies [ 48 , 52 - 54 , 57 , 59 , 68 - 74 , 77 - 80 , 84 , 85 , 88 ], followed by statistical models (9/45, 20%) [ 45 , 49 - 51 , 58 , 81 , 82 ]. The most commonly performed task was recognition (16/45, 36%) [ 52 , 56 , 61 , 63 - 66 , 72 , 73 , 76 , 78 - 80 , 84 , 85 ], followed by forecasting (9/45, 20%) [ 45 , 46 , 49 - 51 , 53 , 54 , 57 , 71 ]. Other tasks performed were event detection, goal-driven optimization, interaction support, and personalization ( Table 2 , Multimedia Appendices 2 and 3 ).…”

“…Although more than half (24/45, 53%) of the AI applications had no action autonomy [ 46 , 48 - 51 , 53 , 54 , 57 , 63 , 66 , 67 , 70 , 73 - 75 , 79 , 81 - 85 , 87 - 89 ], a few reported applications had low (2/21, 10%) [ 55 , 72 ], medium (4/21, 19%) [ 58 , 69 , 71 , 86 ], or high (6/21, 29%) [ 52 , 55 , 60 , 61 , 68 , 76 ] action autonomy ( Table 2 , Multimedia Appendices 2 and 3 ). Nearly three-quarters of all AI systems were intended for clinical care (33/45, 73%) [ 46 , 49 , 51 , 53 - 59 , 61 , 63 - 73 , 78 - 80 , 84 - 89 ], and the majority (18/33, 55%) of these concerned providing support to inform the patient-provider encounter [ 46 , 49 , 51 , 55 , 56 , 61 , 63 , 65 , 67 , 68 , 74 , 78 , 79 , 84 , 85 , 87 , 89 ], followed by diagnosis and prediction-based diagnosis (13/33, 39%) [ …”

“…Most (33/45, 73%) of the articles were from North America and Europe[46,47,[49][50][51][52][53][54][55]57,58,[61][62][63][67][68][69][70][73][74][75][76][77][78][79][80][81][82][83][84][85][86][87], of which most (18/33, 55%) were from the United States[46,47,49-52,54,68,73-77, 79-81,84,87]. The greatest number of AI systems were implemented either in hospital-wide settings (6/45, 13%)[50,55,56,65,74,80] or in radiology (6/45, 13%)[53,56,66,68,73,76]. Most (27/45, 60%) of the studies were authored by a multidisciplinary team[46,47,50-55,58,…”

“…There was a wide range of study designs. Most (24/45, 53%) used a case-study design, including both single-case[46,49,50,52,53,[55][56][57]59,60,[66][67][68]70,74,75,[78][79][80][81][82][83]85,86] or multiple comparative case designs[53,56,78] (Table2and Multimedia Appendix 2).…”