_____________________________________________________________________________ AI in Healthcare
PRIVACY, SECURITY, AND CONFIDENTIALITY ISSUES Health and mental health data are sensitive, and there are concerns about how secure data are when using AI-driven platforms. One of the main questions is the extent to which are there risks of breaches of security, compromising privacy and confidentiality [4]. Patients’/clients’ health records, therapy/ counseling session notes, clinical histories, and behavioral data are required to comply with HIPAA regulations in order to ensure patient/client confidentiality [51]. In datasets, patient data are required to be de-identified; however, sophisticated algorithms are able to re-identify the data [52]. In one study, an algorithm was able to re-identify data from 85% of adults and 70% of children in a physical activity cohort study [53]. Informed Consent At the heart of informed consent and client autonomy is the ability of clinicians to provide sufficient and understandable information to allow individuals to make an informed decision about their care. Patients/clients should be informed about the potential risks and benefits of AI in the delivery of their care. For example, AI algorithms can generate or perpetuate biases, create hallucinations (inaccuracies when the AI misperceives patterns), and introduce cyber-risks (e.g., data breaches) [54; 55; 56]. When a practitioner has employed AI for clinical decision- making without oversight, the patient/client should be made aware of this [56]. Furthermore, they should understand how their clinical/medical/behavioral data may be used beyond their clinical treatment planning [4; 56]. PERPETUATING BIAS AI algorithms are trained on existing datasets, and the valid- ity and reliability of the original data are influenced by the original data collection procedures [57]. If the original data are not representative of certain demographic groups and if it ultimately makes inaccurate predictions, health and mental health disparities, biases, and inequities can be perpetuated [4; 23; 58]. This is known as algorithmic bias, which has been defined as, “the instances when the application of an algo- rithm compounds existing inequities in socioeconomic status, race, ethnic background, religion, gender, disability, or sexual orientation to amplify them and adversely impact inequities in health systems” [59]. Some have advocated for rigorous peer review feedback of AI algorithms to help counteract this potential problem [55]. The following strategies have been proposed to address algorithmic bias [60]: • Including experts who are diversely represented to review AI algorithms • Employing methods to manage situations in which there is not enough available information • Introducing algorithms gradually in order to test their outcomes • Creating mechanisms to collect feedback
LESS THAN OPTIMAL SOLUTION TO REDUCING DISPARITIES Although AI has been proposed as a potential solution to reduce or mitigate health and mental health disparities, this may be an oversimplistic view. Many macro systemic factors will continue to exist and will exacerbate issues related to the digital divide, which continues to affect the general utilization of AI [27]. As of 2024, 80% of U.S. adults had broadband Internet access at home. However, there are significant differences in among racial/ethnic groups. While 83% of White American and 84% of Asian American adults have home broadband access, this reduces to 68% for Black adults and 75% for His- panic adults [61]. It is plausible that AI will actually increase the disparities because the user demographics will be reflected back to users [27]. LACK OF TRANSPARENCY, ACCOUNTABILITY, AND POLICIES/REGULATIONS AI algorithms have been compared to black boxes, because it is not always clear how the algorithm was designed, developed, tested, and validated [58]. Private commercial companies are often wary of releasing too much information or risking their intellectual property [62]. This is counter to the transparency regarding the validity and reliability and safety and risks of AI-derived clinical information health and mental health professionals require in order to make sound clinical decisions for their patients/clients. Similarly, patients/clients should understand the nature of the information in order to make informed decisions about their care [58]. In a 2024 study assess- ing public documentation on 14 AI-based radiology products, researchers employed a self-designed instrument to measure the transparency of the AI products in different domains, includ- ing ethics, safety and risks, training of the AI algorithm, and performance limitations [58]. Overall, they found very little publicly available information about the products, particularly on safety and risks. Most of the products did not disclose how the AI was trained to evaluate for algorithmic biases. Unfor- tunately, because AI technology has grown and changed so rapidly, regulations and policies have not kept pace. RISKS/SAFETY, PROFESSIONAL LIABILITY, AND COMPETENCE Data error is a challenge to the integrity of AI applications. If there are any algorithmic inaccuracies and/or biases, false clinical decisions could be made and result in harm to patients/ clients. This in turn can impact professional liability [4]. Hal- lucinations are relatedly common with most AI applications; these events occur when AI software misperceives patterns that are do not exist, and it then produces nonsensical, inaccurate, and fabricated data and outcomes [63]. Ultimately, hallucina- tions can result in misdiagnoses, medical errors, inaccuracies in medical documentation, and data inaccuracies, leading to faulty and potentially harmful conclusions [54]. The follow- ing strategies may be incorporated in order to build trust and mitigate the impact of AI hallucinations [54]:
41
EliteLearning.com/Physician
Powered by FlippingBook