Role of clinical judgement in AI powered healthcare

Deep learning and machine learning have infested every known field in last couple of years. Healthcare has not remained immune to it either. However, there is much more to healthcare than improve AUC and reducing errors. The cost can be too high. This tutorial discusses trends of using AI in healthcare, from automating electronic health records and using it to predicting patientcare, radiology, ophthalmology to genomics, other omics and eventually personalized medicine. Once the AI powered healthcare starts yielding results "better than" doctors, then the clinical deployment becomes the next critical stage. Clinical judgement involves clinical research, experience and other supporting sciences. The presentation discusses from a generic ML/DL workflow, how unintentional small errors in each step can lead to spurious predictions.

This tutorial will trace the journey of possibilities for deep learning in healthcare and how an integrated, holistic use will assist doctors and hospitals in providing targeted healthcare.

 
 

Outline/Structure of the Tutorial

  1. Introduction and organization of talk (5 min)
  2. Use of AI in healthcare (5 min)
  3. A typical healthcare AI system and sources of artefacts (5 min)
    1. Missing information.
    2. Often datasets are small and jump-started with pre-trained models with their own bias
    3. Causality is always overlooked mainly because the data is observational.
    4. Pitfalls in using clinical data (patient monitor data)
    5. Model overfitting
    6. Model metrics
    7. Bias
  4. Clinical Judgement examples in AI systems (15 min)
    1. Examples from radiology and ophthalmology - can the model detect the disease itself or treatment artefacts or both? These are the "label leakage" and "label misclassification" problems. What cliical challenges this poses?
    2. Extending the radiology and ophthalmology examples - as data is randomly split into training and testing during training, what if the same patient's data becomes part of both train and test datasets? What clinical challenges it poses?
    3. Example from ophthalmology and EHR data - can the model be confounded? What are implications on clinical judgment?
    4. Examples from EHR datasets - What happens if we add data after we find issues in stratified analysis? Does just adding data in strata improve the model performance? Bias in the model comes from variety of factors and simple fixes may not fix the model performance.
    5. Clinical deployment of a model over long period of time - The model has not changed over time. same version is used. But database is enriched over time and we find the model starts failing after a few months? Why? This is the "dataset shift" problem.
  5. Concluding remarks in deploying healthcare solutions (10 min)
    1. Patient safety
    2. Patient-doctor relationship
    3. Public acceptance and trust
    4. Accountability for decisions
    5. Bias, equality and fairness
  6. Q & A (5 min)

Learning Outcome

  1. Understanding of how deep learning can be applied typically in healthcare
  2. Understanding role of clinical judgement in AI systems
  3. Awareness of pitfalls and limitations of deep learning in healthcare.
  4. Yet AI alone can help personalized medicine - but it has to be used as assisting tool.

Target Audience

AI developers, Healthcare professionals

Prerequisites for Attendees

  • No specific exposure
  • The lecture will introduce concepts in both AI and Omics before the deep dive
  • Organized more as a popular talk
schedule Submitted 1 year ago

Public Feedback


    • Liked Anoop Kulkarni
      keyboard_arrow_down

      Anoop Kulkarni / Akshara Kulkarni - Deep genomics of type 2 Diabetes

      20 Mins
      Talk
      Advanced

      Type 2 diabetes mellitus has been at the forefront of human diseases and phenotypes studied by
      new genetic analyses. Thanks to genome-wide association studies, we have made substantial
      progress in elucidating the genetic basis of type 2 diabetes.

      This tutorial summarizes the concept, history, and recent discoveries produced by genome-wide association studies for type 2 diabetes and glycemic traits, with a focus on the key notions we have gleaned from these efforts. Genome-wide association findings have illustrated novel pathways, pointed toward fundamental biology, confirmed prior epidemiological observations, and provided possible targets for pharmacotherapy and pharmacogenetic clinical trials.

      Deep neural networks can be trained to detect Type 2 diabetes using a few biochemical and anthropogenic measurements. Such a network can then be used to predict onset of diabetes in other patients with access to these parameters. The tutorial will discuss a few other applications of deep learning in genomics of type 2 diabetes.