Responsible AI in Healthcare

Siddhant Khare • September 16, 2021
Share This Article


Introduction to Responsible AI

As Artificial Intelligence continues to garner the rightly deserved limelight, it also takes giant strides to create a positive impact in the world. However, it is imperative to keep in mind that AI is not completely error-free, and it cannot replace human judgment.

Thus, to enable AI to achieve its full potential we need tools in our arsenal that can tackle the existing chinks in AI’s armor, especially when it comes to taking high-stake business decisions. Responsible AI allows us to look at the developed AI systems through the much-needed lens of responsibility, leveraging AI with trust, fairness, and security.

Responsible AI in Healthcare

Why is it important?

Traditionally, around 80% of the decision-making in healthcare has been relying heavily on heuristics and rule-based systems. Recently, this trend has been shifting towards AI/ML systems (in areas such as radiology, imaging, etc.) since these models can consider thousands of features, learn complex patterns, and deliver accurate predictions.

Yet, it seems to be a justifiable reason for NOT holistically adopting AI/ML systems for healthcare decisions, especially for patient care: Whom to hold “Responsible” for the predictions made by the AI/ML solutions?

For this very reason, it has become imperative to incorporate “Responsibility” into the predictive models, which can be achieved through 3 fundamental principles of Responsible AI.

Principles of Responsible AI in Healthcare

a) Justness

While there are massive benefits of using AI in healthcare, there looms a larger risk of decision making based on the bias introduced in the AI/ML solution through Data and Algorithms. Therefore, detecting & mitigating biasness is of utmost importance and can be achieved throughout the development cycle of AI/ML solutions using techniques such as Pre-processing, In-Processing, and Post-Processing. While the potential human costs of biased AI should remain a primary concern, the lack of identification & mitigation of biasness could also lead to legal and financial implications that inappropriate, inaccurate, or discriminatory algorithms can have for organizations. Hence, Identifying, and mitigating biases in AI models is the 1st step to foster fair data-driven Healthcare decisions.

Tools/Techniques: IBM AI Explainability, Microsoft’s Fairlearn & AWS’s Fair Bayesian Optimization

b) Explainability

It is a no-brainer to state that in Healthcare, Black box AI models create doubts and confusion, and therefore alleviation of this reasoning uncertainty of AI models can be done by adding a dimension of explanation or transparency to the models. The explanation of models can be classified into 2 categories:
1) Post hoc explanations for Black-box models which could either be Local or Global
2) Glass Box Models which by themselves are inherently interpretable

Explainable AI provides a mechanism to build trust in AI Model predictions (for example, pointing out the pixels in a chest X-Ray explaining the emergence of a disease) without compromising on the accuracy of the models and ensuring that legal & regulatory measures are being satisfied.

Tools/Techniques: SHAP, LIME, Partial Dependence Analysis & Individual Condition Expectation Analysis

c) Privacy-Preserving

Availability of humongous amounts of data (especially in healthcare) has been one of the biggest reasons for the incredible breakthroughs that we are witnessing in Artificial Intelligence, but, as it turns out, “Data” can prove to be a double-edged sword if not handled with care. Hence, healthcare data security and privacy in AI are becoming increasingly the topmost priorities to safeguard patient’s data and other confidential information.

 The methodology for achieving Privacy in AI is through the utilization of advanced properties & techniques such as Differential Privacy & Federated Learning. In other words, they can be summarized as follows:

 1) Differential privacy: Protecting the data before it gets consumed by the AI model

2) Federated Learning: Building protection inside the AI model

 Tools/Techniques: IBM’s Diffpriv, Microsoft’s WhiteNoise, Google’s Tensorflow FL 

Key Benefits of Responsible AI in Healthcare

Incorporating principles of Responsible AI into the existing AI/ML models in the healthcare domain promotes:

a) Identification & mitigation of biasness in the AI/ML solutions
b) Faith in the future(predictions) without compromising on the accuracy of the models reduces the risk for disparate impacts on vulnerable populations and limits the risk of legal and financial implications
c) Decisiveness in the present
d) Safe & Secure environment for the AI model ecosystem

Let’s create something amazing together!

Contact us Next
Latest Blog
LinkedIn Instagram Facebook Twitter