Transforming HealthCare: AI’s Role in overcoming industry challenges

Pic: https://www.globusmedical.com/musculoskeletal-solutions/excelsiustechnology/excelsiusgps/

AI/ML in healthcare and life sciences have an transformational impact across the industry. It can assist the healthcare professionals in providing the efficient care, find a cure for incurable diseases, building a genomic & clinical NER application, applying clinical trials and PharmacoVigilance (PV) , drug discovery design, pharmaceutical supply chain & sales and clinical NER application.

However, there are also certain challenges in implementing AI models in this industry. Major challenges.

Confidential and private Data

Regulations

Black Box Modeling Approach from complex Deep Learning models

Highest level of security

While there are so many challenges ahead of us, there are a quite a good ML stories in healthcare and life sciences in the recent past.

We have seen companies like Phillips, GE have revolutionized the medical imaging through Computer Vision products

1. Adaptive Intelligence, Radiology workflow optimization which automates routine tasks and let radiologists foucs on complex use cases.
2. Illumeo which uses ML to analyze medical images and patient data to develop personalized treatment plans based upon the individual characteristics
3. Dosewise which actively reduces the radiation exposure without compromising on image quality and side effects for patients.
4. Image Guided Therapy gives minimally invasive procedures thorugh real time imaging feedback.
5. Edison which uses medical images and patient data to enable faster and accurate diagnosis across imaging apps including CT, MRI and ultrasound.

Lets deep dive into the challenges and see how we can overcome them and help us lead a better life together without comprising on the data.

Challenge 1: Confidentiality and Privacy

There are severe fines and penalties to the companies who do not comply with the rules set by GDPR or HITECH. Total fines were approximately $80 million as per one of the repots.

Challenge 2: Regualtions

Regulations like HIPAA which works for protection of Protected Health Information (PHI) , HITECH for electronic health Record , GDPR for Personably Identifiable Information (PII), GxP for ensuring safety and quality.

Challenge 3: Generalizability and Reproducibility

Model interpretability can be challenging as they become more complex. Ethical considerations around privacy, confidential and informed consent can impact data collection and the ability to reproduce and generalize ML models.

Challenge 4: Transparencies and Human Bias

A biased model in healthcare industry can have serious effects to the society where health benefits are supposed to be equal for all the ethnicity, race and genders. For example, an algorithm used to detect breast cancer from mammograms was found to be less accurate for women with denser breast tissue. Socio economic bias where patients would benefit from additional care was found to be biased agianst patients who lived in low-income areas.

Few Options to solve these challenges

Agreed that there are mounting challenges and it can never replace a trained medical professional who has relevant practical experience and knwoledge gained through face to face interactions. It can only help these professionals in taking a tangible decision from the list of options provided by ML. Also to automate few of the mundane tasks like searching medical history info, case studies, research papers to expedite their decision making process.Few of technigques to consider in your day to day AI workloads as you overcome these challenges above.

Amazon Comprehend

1. Encrypting and anonymization of data — By using secret key and data masking techniques to remove the sensitive or PII data. Generating synthetic data versions to replicate the original data set.
2. Bias and Explainability of decisions — Using LIME to provide local explanation of black box models and check how permutations of future values impact on overall model. SHAP that captures feature contribution. Permutation importance can also be used.

Metrics: Class Imbalance, Difference in Proportions label, KL divergence, Conditional Demographic Disparity in Labels, Recall Difference, Difference in Positive Proportions in Labels, Difference in Label Rates, JS Divergence.

3. To comply with compliances, building AI pipelines which has versioning enabled and able to reproduce the results or predictions based on dataset.

Logging and versioning of the AI pipelines.

4. Human in the Loop to check if there are any dips in performance and model degradation with SME.

Active Learning Pipelines, Regular Feedback capturing mechanism using AIOps, Responsible AI to check usage policies

For a detailed walkthrough on how to implement the active AI pipelines. I recommend the articles/books below.

1. https://www.amazon.com/Applied-Machine-Learning-Healthcare-Sciences/dp/1804610216
2. https://www.eecs.mit.edu/research/explore-all-research-areas/ml-and-healthcare/
3. https://www.hipaajournal.com/healthcare-data-breach-statistics/
4. https://www.reuters.com/legal/litigation/three-us-data-breaches-show-varied-healthcare-exposure-risks-2023-02-06/

In the next article, I will be coming up with what lies in the future of health care and life science industry and trends which we will be seeing in the next decade to make our lives better.