Page 163 - Kaleidoscope Academic Conference Proceedings 2024
P. 163

Innovation and Digital Transformation for a Sustainable World




                                                              potential of our approach to bridge the gap between general-
            Can I take   Ibuprofen is a nonsteroidal anti-inflammatory   purpose  language  models  and  domain-specific  health
            ibuprofen    drug (NSAID) used to treat pain and   applications.
            while on     inflammation, while lisinopril is an ACE
            lisinopril for   inhibitor used to treat hypertension and other   The representative examples in Table 2 illustrate the system's
            my blood                                          ability to provide personalized, actionable recommendations
            pressure?    cardiovascular conditions. Taking ibuprofen
                         while on lisinopril can potentially decrease   based  on  users'  specific  health  contexts  and  needs.  By
                         the effectiveness of lisinopril and increase   leveraging  knowledge  retrieval,  the  system  can  tailor  its
                         your risk of kidney problems. If you need   outputs to individual users while maintaining alignment with
                         pain relief, acetaminophen is generally   established clinical guidelines and best practices. This level
                         considered a safer option. However, it's   of  personalization  is  critical  for  engaging  users  and
                         always best to consult with your doctor or   promoting  behavior  change,  as  generic,  one-size-fits-all
                         pharmacist before taking any new     health advice is often less effective. However, our work also
                                                              highlights important limitations and challenges that need to
                         medications to avoid potential drug   be addressed. One key issue is the potential for biased or
                         interactions. They can recommend the most   inconsistent outputs, particularly when dealing with complex
                         appropriate treatment options based on your   or ambiguous health queries. While our retrieval-augmented
                         individual health needs and medical history.
                                                              generation approach helps mitigate this risk by grounding
                                                              outputs in verified knowledge, there may still be cases where
                                                              the model generates inappropriate or misleading responses.
           Table  2  illustrates  two  representative  user  interactions,   Developing  more  robust  methods  for  controlling  and
           showing how the RAG incorporates retrieved knowledge to   aligning model outputs, such as adversarial training, value
           personalize  and  strengthen  its  responses  compared  to  a   learning,  or  human-in-the-loop  oversight,  is  an  important
           baseline without augmentation. For example, when the user   direction for future work [42], [43]. Another challenge is the
           asks  about  lifestyle  changes  for  hypertension,  the  system   need  to  continuously  monitor  and  update  the  system's
           draws  on  authoritative  guidelines  like  DASH  to  suggest   knowledge  bases  to  keep  pace  with  the  rapidly  evolving
           tailored diet and exercise tips. The drug interaction query   health  landscape.  As  new  research  findings,  treatment
           triggers a safety warning and recommendation to consult a   guidelines, and public health recommendations emerge, it is
           doctor,  based  on  structured  information  from  a  medical   critical that the system's underlying knowledge is updated
           database.  These  examples  highlight  how  our  framework   accordingly.  This  requires  ongoing  curation  and
           enables more informed, actionable, and context-aware health   maintenance efforts, as well as mechanisms for detecting and
           advice  by  dynamically  integrating  relevant  domain   mitigating  potential  inconsistencies  or  conflicts  between
           knowledge  into  the  generative  process.  The  personalized   different  knowledge  sources.    Privacy  and  security
           outputs also establish a meaningful user dialogue, while the   considerations  are  also  paramount  when  deploying  AI
           retrieved facts help maintain clinical validity.   systems in the health domain. While our approach does not
                                                              directly use or store personal health data for model training
                            5.  DISCUSSION                    or inference, there may still be risks of sensitive information
                                                              being  inadvertently  revealed  through  user  interactions.
           Our results demonstrate the potential of generative AI and   Techniques  for  privacy-preserving  AI,  such  as  federated
           knowledge  retrieval  to  enable  personalized  digital  health   learning, differential privacy, and homomorphic encryption,
           services.  By  combining  the  strengths  of  large  language   could help mitigate these risks and ensure compliance with
           models, which can engage in fluent, contextual interactions,   data protection regulations [44]. It is important to recognize
           with curated health knowledge bases, which provide verified,   that our system is intended to supplement, rather than replace,
           domain-specific  information,  our  proposed  system  can   human healthcare providers. While generative models can
           provide users with relevant, reliable, and actionable health   provide valuable information and support, they should not be
           support.  The automated evaluation results suggest that our   used  for  definitive  diagnosis,  treatment  planning,  or
           system can generate high-quality, accurate responses to user   emergency response. Ensuring appropriate use and setting
           health  queries.  The  low  perplexity  and  high  BLEU  and   realistic expectations for both users and providers is critical
           ROUGE  scores  indicate  that  the  generated  text  is  fluent,   for the safe and effective deployment of AI in healthcare.
           coherent,  and  aligned  with  human-written  references.  The   There  are  also  vital  considerations  around  responsible
           factual  accuracy  of  92%  is  particularly  encouraging,  as  it   development   practices,   model   interpretability,   and
           shows  that  the  system's  outputs  are  grounded  in  verified   stakeholder   involvement   that   require   ongoing
           health information. This is a critical consideration for any AI   multidisciplinary collaboration to address. Domain experts
           system deployed in the health domain, where inaccurate or   such as clinicians, patient advocates, ethicists, and regulators
           misleading  information  could  have  serious  consequences.   should be engaged throughout the research and development
           The user study results further validate the system's utility and   lifecycle to align system capabilities with real-world needs,
           usability.  Both  lay  users  and  healthcare  professionals   values,  and  constraints.  This  includes  proactive  risk
           reported  high  satisfaction  with  the  generated  responses'   assessment and mitigation strategies around safety, fairness,
           relevance, usefulness, and clarity. The positive ratings from   transparency,  and  accountability.  Policymakers  and  health
           medical experts also suggest  that the system's outputs are   system  leaders  will  also  need  to  establish  governance
           clinically valid and complete. These findings underscore the




                                                          – 119 –
   158   159   160   161   162   163   164   165   166   167   168