Patient - First Health With Generative AI

Patient - First Health With Generative AI

• 1.
  Patient-First Health with GenerativeAI
• 2.
  Insights generators: GenerativeAI enables real-time analysis of both structured datasets (such as health data) and unstructured datasets (such as physician notes), offering significant promise as a “co-pilot” to help healthcare providers quickly analyse information and interpret results. These tools also can help accelerate data integration across datasets to help doctors, nurses and health systems gain faster, more accurate insights. Three broad categories of generative AI use cases for patient engagement 1 Action drivers: Traditional chatbots have frequently frustrated both patients and providers with incomplete, confusing or irrelevant information that fails to convert insights into action. Generative AI can augment providers by offering more intuitive, human-like conversations with patients and caregivers that can help inform care decisions and spur action. Productivity boosters: With minimal prompting, generative AI models can create images and text sufficiently human-like as to automate many of today’s manual tasks – everything from transcribing doctor-patient visits and drafting emails to summarizing clinical studies and dispensing general health information. Generative AI can also be used to help create a more robust data foundation by automating data management and extracting insights from unstructured data, a task that was extremely time-consuming, cumbersome and inefficient using earlier AI techniques.
• 3.
  How and whygenerative AI will revolutionize the way patients engage with their health Assist in monitoring, side effect management, adherence and keeping patients on treatment Disease management interventions Co-pilots for patient triage Patient care navigation in the first mile to help understand where to direct patients for diagnosis or treatment Lifestyle support, early disease education and even prenatal care in LMICs*, with a focus on consumers and caregivers Health education assistance *Low- and middle-income countries 2
• 4.
  Three barriers tosafe, effective patient-facing generative AI 3 As with many novel technologies, there is a risk that generative AI exacerbates existing disparities. Multilateral, cross-border partnerships will be key to accelerating generative AI’s impact outside high- income countries. Two barriers stand out: Language, Resources Scalability across regions and contexts Holes in the data foundation While many organizations have made progress in building their own data management infrastructure and augmenting it with data consortia and even federated learning, additional work is needed to ensure models are trained on sufficiently robust data to produce outputs that are fit enough for safe and reliable patient use. Two data issues predominate: Biased outputs, Patient privacy Interviewees singled out trust as the most formidable barrier to widespread adoption of generative AI for patients, with one blanket issue rising above all others: inaccuracy. Mistrust
• 5.
  Build trust throughempathy and interactivity. To instil trust in generative AI models, providers, payers, biopharma and tech companies must prioritize fine-tuning models to make them more empathetic – including by having doctors “test” responses to improve outputs. When these models have been trained on healthcare domain-specific data, and users can engage them in a dialogue to evaluate their outputs, then patient trust and engagement should follow. Mitigate against bias. Providers, payers, government and biopharma all possess rich consumer and patient datasets, but they sit in disconnected data layers. Given that it’s impossible to fully correct for bias inherent in models’ training data, mitigating output bias requires creating a richer data layer by connecting the data ecosystem and then measuring response bias on an ongoing basis – especially in underserved populations. Keep humans in the loop. Generative AI’s uncanny ability to mimic human-like text creates a false sense of security. Yet the technology is fallible. All stakeholders must help educate patients and providers on generative AI’s limitations, especially its potential for hallucination. At the same time, providers must work to integrate patientfacing tools into existing clinical workflows and implement processes to review recommendations and intervene accordingly, especially in high-risk discussions and with patients with severe disease. Plan to scale across contexts. Running large models is extremely energy-consumptive, so technology companies must work to develop and deploy more cost-efficient means to run them. At the same time, governments, NGOs and connectors (defined here as advocacy and patient groups, distributors, group purchasing organizations, investors and other organizations in healthcare) must also develop flexible deployment models that recognize the varying needs and exigencies across regions, cultures and contexts. With generative AI and healthcare, one size does not (and cannot) fit all. 1. 2. 3. 4. To make generative AI for patients a reality, healthcare stakeholders must take these four steps The recommendations below are intended as a generative AI-specific action plan for health and healthcare players to begin addressing barriers to the widespread adoption of patient-facing generative AI. Source: World Economic Forum 4