Dr. Erkan Hassan

Monthly Writings

Evaluations and reviews of the latest in the field.

Making Artificial Intelligence (AI) Bullet Proof

SUMMARY:

  • AI’s prevalence within healthcare has the potential to transform patient care and outcomes.

  • Due to the persistence of common challenges, deployment of AI in clinical practice remains limited.

  • Solutions in the development, implementation and post-implementation can help to make AI solutions more trustworthy.

REVIEW

AI Objectives and Use Cases

OBJECTIVES

  • Improved diagnostic effectiveness

  • Decreased healthcare costs

  • Improved information sharing

USE CASES

  • Diagnostic predictions with early symptom analysis.

  • Virtual health assistants

  • Targeted therapy recommendations

  • Drug Discovery

AI Challenges

TOP CHALLENGES FOR AI

  • Data Quality & Accessibility

  1. Volume of data ≠ quality & availability

  2. Healthcare data is fragmented across various systems contributing to inaccuracies and inconsistencies

  3. May provide inaccurate insights and recommendations adversely affecting performance and reliability.

  • Implementation Costs

    1. Requires significant investment

    • Technical infrastructure

    • Regulatory compliance

    • Staff training

  • Acceptance of AI in Healthcare

  1. The AI Chasm: discrepancies in clinical efficacy and usefulness between predicted model in development dataset and real world data

  2. Lack of trust in results

  • Lack of AI Performance Evaluation

  1. There is a need to assess AI performance before widespread implementation across a health system

  2. Most AI studies are retrospective and based on patient medical record data.

  3. Need reliable measurements of accuracy and reliability

  4. This ongoing gap is a major challenge in determining how well the AI solution meets:

    • clinical goals,

    • integration with existing workflows and processes

    • patient outcomes

    • impact costs

  • Regulatory Issues

  1. Keeping up with the ever changing regulatory environment across the globe

  • Bias

  1. AI algorithms may have inherent bias in economic, social, gender and race

Solutions to Strengthen AI

  • Data Quality

  1. Missing data; implausible data; outlier data

    • AI should identify these types of inputs

    • Allow model to discard or fix

    • Infer useful information associated with faulty inputs

    2. AI should be developed to be resilient to data variability over time, across settings and sites

    3. Data quality should be maintained not only during development, but in production as well.

  • AI Recommendations

  1. Strengthening AI models become critical in output predictions of disease and care.

  2. Uncertainty is intrinsic to decision making.

    • Minimizing uncertainty or informing the user of the amount of uncertainty improves the trustworthiness of the model

  3. Transparency of why an AI system comes to a final recommendation and provides a specific output from varying inputs improves interpretability.

  4. Ability to compare current clinical scenario to similar past cases and their respective outputs would be helpful

  • With Routine AI Use

  1. Once AI solution is in real world routine use, changes in a variety of areas may occur impacting output recommendations:

    • Additional data

    • Clinical knowledge

    • Changes in target population

    • User requirements

    2. Monitoring, benchmarking and reporting performance on dataset shifts becomes critical

    3. AI systems which can adapt to new knowledge over time, require greater transparency, especially when moved to a new clinical setting

  • Regulatory

  1. As rules on transparency, liability and legal frameworks evolve, AI systems need to be able to be:

  2. Updated automatically

  3. Notify users of changed situations

  4. How to implement and notify of new rules is left to system designers

  • Bias

  1. Identify underrepresented sub populations or where an imbalance exists

  2. Re-weight AI training to mitigate the discrimination

  3. Remove variables that do not contribute to output performance

  4. Make changes to model in pre-development; during training of the solution and post-training.

CONCLUSIONS:

  • There are a number of challenges currently present which preclude the widespread use of AI in clinical care.

  • Despite these challenges AI will be transformative within health care.

  • Developing or evaluating AI systems should include the complete assessment of the areas outlined above.

Erkan Hassan