What are the potential risks and pitfalls of relying too much on AI in HIT?

I would be interested in how we leverage AI to make less biased and/or hasty diagnoses and treatment recommendations. Human providers biases, and time constraints can lead to poor history taking, unfounded assumptions, and patient distrust, and subsequent failures to disclose pertinent details. Might patients eventually trust AI more?

One such risk is the possibility of inaccurate or biased predictions or decisions made by the AI algorithms, leading to incorrect diagnoses or treatment plans. Additionally, an overreliance on AI may result in the degradation of human clinical skills and judgment, potentially diminishing the doctor-patient relationship and the quality of care provided. Furthermore, the reliance on AI may raise concerns regarding data privacy and security, as massive amounts of sensitive patient information are processed and stored by AI systems.

While the use of AI in healthcare shows immense promise, I have seen firsthand the ethical and legal complexities it presents. In one instance, we implemented an AI system that analyzed patients' electronic health records to predict specific health outcomes. Despite the tool's effectiveness, it raised concerns regarding privacy and the consent of patients. We had to ensure that the tool's use complied with HIPAA and GDPR, among other regulations, and was transparent in its decision-making process. Balancing the potential of AI with respect for patient rights is a continuous and necessary dialogue in the healthcare community.

With EMR systems updating automatically and monstrous amounts of patient data accumulated esp for chronically ill patients - some accurate, some unreliable, some blatantly inaccurate - how can we ensure that AI “thinking” is based on quality information? If clinicians document the diagnosis, symptoms, meds, etc incorrectly-especially in rare disease patients- and the chart contains extraneous information, the AI will not produce quality output. Correcting the chart isn’t easy, as one hospitalization might mean hundreds of pages of data, an insurmountable task for well patients let alone the chronically ill. Multiply this problem by years and the charts of chronically ill patients present a huge barrier to effective use of AI.

For widespread adoption of AI in HIT, we FIRST need adoption of Good Machine Learning Practices (GMLP) -that are very similar to good software engineering practices or quality system practices. These GMLP should describe a set of AI/ML best practices (e.g., data management, feature extraction, training, interpretability, evaluation and documentation) that are then adopted and standardized by all healthcare companies/products adopting AI.

In my practice, one significant challenge was managing the quality and diversity of data required for our AI systems. For example, we once worked on a machine learning model to predict disease progression, but the model struggled because the quality and completeness of the historical data varied significantly. Also, issues like overfitting or underfitting of models presented real problems. These experiences emphasized that we need robust, quality data and meticulously trained models for successful AI implementation in HIT.

Some of the technical challenges of operationalizing AI in HIT can be mitigated by: 1. Concept normalization: critical process in AI that involves mapping diverse lexical expressions and clinical terms to a standardized concept or reference terminology allowing interoperability 2. Deep learning: computational method utilizing artificial neural networks (ANNs) with multiple layers of interconnected nodes allowing discovery of intricate patterns that would otherwise remain elusive to detect using conventional statistical approaches 3. Federated learning: a novel paradigm in AI/ML that enables collaborative model development across multiple healthcare organizations, while preserving the privacy and security of sensitive patient data.

AI in healthcare IT introduces technical and operational challenges. Address issues like data quality, bias, and availability. Tackle challenges of training, validating, and updating AI models. Ensure reliability, scalability, and interoperability of AI systems. Monitor and improve performance, testing for accuracy, robustness, and safety.

Prioritize interpretability and explainability: Select or develop AI models that provide transparent explanations for their decisions, aiding in error analysis and building trust. Conduct thorough testing and validation: Implement a comprehensive testing strategy across diverse scenarios to assess the accuracy, robustness, and safety of AI systems before deployment. Cultivate interdisciplinary collaboration: Foster strong partnerships between healthcare professionals, IT experts, and data scientists to address the complex challenges of AI implementation and ensure optimal outcomes.

AI systems, particularly deep learning models, can be complex and lack transparency. This "black box" nature makes it difficult to understand how decisions are made, complicating troubleshooting and trust. Integrating AI with existing HIT infrastructure can be technically challenging and costly. Compatibility issues and the need for extensive customization can hinder smooth implementation.

Don't implement AI solutions unless you're confident they add value to your users. Also, don't treat your users as guinea pigs because you will lose their trust. Implement only proof-of-concept solutions. This is especially important when you develop mental health solutions, and your users rely on your product to manage depression, anxiety, and other mental issues. While AI might offer support, in many cases, it should not be the primary option for your users.

From colon polyp detection to chatbots, AI has taken healthcare by storm. It's important to recognize, that these are just adjunct technologies for now. The provider still is the one deciding whether the AI is correct. Its too early to solely rely on AI findings, without interpretation.

We are entering a new era with abundance of information but a scarcity of time and human connection in Healthcare. Healthcare is more than just processing information and associating words with concepts; It is ascribing meaning to those concepts and while connecting with patients as a trusted partner to build healthier lives. AI can play a key role in enhancing telehealth services, diagnosis and/or efficient information exchange between patients and healthcare providers but there is no substitute for the human connection and clinical judgement....YET!!

In my journey, implementing AI in HIT wasn't solely a technical challenge but also a human one. For instance, integrating AI applications into clinical workflows required significant changes in how physicians worked. Some staff were initially resistant to these changes and skeptical of AI's ability to provide clinical insights. To overcome this, we had to foster a learning culture and ensure all team members felt engaged with and understood the technology. Our focus was on striking a balance between leveraging AI benefits and ensuring the comfort and trust of our healthcare providers.

AI, as it stands, requires additional testing and training in order to be safe and effective for rare disease patients. AI can be of immense value to Zebras in facilitating rare diagnoses, but AI can also give human doctors a sense of false confidence if it misses the diagnosis and says the patient is fine. Humans must recognize that AI isn’t infallible and AI errors can endanger human life. Example-I recently gave AI a scenario of acute abdominal pain in a patient w/ clinical adrenal insufficiency. The AI suggested mental health interventions, missing the correct diagnosis of potentially life threatening Adrenal Crisis. It wasn’t until I mentioned Adrenal Crisis that the AI acknowledged bias and the need for more training on rare diseases.

Consider AI in HIT as a powerful assistant rather than a replacement. Embrace AI as a tool to augment human capabilities, not overshadow them. For instance, in diagnostics, use AI to assist radiologists in identifying potential issues, but maintain the critical human touch in patient communication and empathy. This approach ensures that AI complements healthcare professionals, leading to more accurate diagnoses and better patient care. By striking a balance between AI and human expertise, you harness the strengths of both for improved healthcare outcomes while minimizing the risk of overreliance on AI.

AI relies on data science, so it's crucial to ensure you have an adequate amount of data to support your statistical decisions. This becomes especially vital in high-risk AI applications, such as disease diagnosis, where you use natural language descriptions of medical images or patient records to aid radiologists and doctors in their diagnoses. In such cases, there is no room for error since human lives are at stake.

Exciting times ahead in the healthcare sector as AI becomes an integral part of HIT. As a founder in the life science industry and AI enthusiast, I emphasize the urgency of marrying AI's potential with human expertise. We need a human-in-the-loop framework to maintain human judgment and empathy at healthcare's forefront. Equipping professionals with AI literacy and establishing ethical protocols is crucial to protect the patient's data and the caregiver alike. We have to set the goal to create an inclusive, trustworthy and accessible Healthcare system assisted by Ai, so we can focus on what matters: human connections.

In my experience, we need to make sure that, amid the AI revolution, the human element does not get overlooked. A balanced integration of AI that leverages a human-in-the-loop framework is a potential approach. This approach will ensure that human judgement stays at the core of healthcare, leveraging our experience and intuition. It will also allow for the optimisation of AI algorithms through continuous feedback and improvement. This leads to the obvious need to equip healthcare professionals with AI literacy. In addition to this, protocols for ethical use must be established. A combination of these will ensure we not only leverage the potential of AI but also keep the indispensable human touch.

As a physician and AI enthusiast, I see an exciting future for AI in HIT. However, we must acknowledge and prepare for potential challenges. For instance, the need for high-quality, diverse data will increase as we move towards personalized medicine and preventive care. We must also ensure our AI strategies align with emerging technologies like blockchain for data security or quantum computing for faster data processing. The ultimate aim should be health equity, ensuring that the advantages of AI in HIT are accessible to all, irrespective of their geographical location or socioeconomic status.