top of page

AI Can Only Drive Health Equity With Diverse Datasets

Mainul Mondal is the Founder, Chief Executive Officer, and Member of the Board of Directors of Ellipsis Health.

This is a pivotal moment in healthcare AI. For several years, AI applications have been carefully designed and prototyped to use the vast amount of existing health data for novel purposes, driving better insights, earlier detection of illnesses and, ultimately, more accurate healthcare. All of a sudden, the breakthrough of the public availability of generative AI tools has created an explosion of interest and investment in the space. Now that the public and the healthcare sector are catching up on the last decade of advancements in machine learning, we’re presented with a momentous opportunity to improve health across the board. AI is moving faster than any emerging technology ever has and debates on its accuracy and pitfalls are moving at the same rate. Mistakes have dire consequences in healthcare and with so much talk of possibilities and potential, I believe it’s important to return to first principles:

  • Healthcare technology should improve outcomes for all individuals; any AI solution used in healthcare must be capable of generating or incorporating accurate diagnoses of health status across diverse populations.

  • Using technology like AI to measure, evaluate and intervene in mental health care may be the only way to bridge the gap in care capacity in a scalable way.

  • The fundamental truth is that data must be both large in scale and diverse in representation to train truly equitable and effective algorithms across different populations.

This perspective reminds us that any building is only as strong as its foundation; and to create a meaningful, lasting impact on health equity, we must ensure that our tools are built with the right data.

Not all the challenges in overcoming health inequities will be solved by AI, but without it, we might continue to rely on methods of diagnosis and resource allocation that are susceptible to human error. To improve health equity across the board, new technologies must be developed to bridge the gap, but this must be done responsibly, taking into consideration the diversity and the social determinants of health of the populations healthcare companies serve.

The fundamental advantage of AI applications in healthcare is that a model trained on tens of thousands of labeled data points is capable of operating in a way that can meet today’s problems of scale in healthcare while also improving accuracy, accessibility and reliability. Models being tested both in studies and real-world applications today have been trained on the data of more patients than a human radiologist, behavioral health specialist or general practitioner can assess and treat during their entire career; and can produce a more objective measurement of an individual’s health by relying on that data.

Importantly, researchers and engineers can make it possible for a model to be trained on a more diverse range of patients than a human doctor. These models will not replace practitioners, but they can help them make more informed, sensitive decisions about the people they care for and save them valuable time so they can focus their expertise where it is most needed.

For example, my company developed a combined acoustic and semantic language model that identifies, measures and monitors the severity of depression and anxiety using voice. We are one team among many pioneering scientists, engineers and healthcare specialists whose innovations, when built with representation and health equity as a foundational principle, will set a completely new standard of care.

If the transformation of healthcare is truly our goal, it is of the utmost importance that healthcare innovators design their solutions from the ground up with health equity in mind. Companies should be transparent about the datasets and methodologies of their training models, and take measures to ensure that these models are unbiased. The consequences of ignoring diversity will lead to more tangible negative impacts on the people we hope to serve, and we’ll miss an incredible opportunity to truly change our field.

With this in mind, any founder or business leader who is developing an AI-powered tool for healthcare should establish a set of operational procedures necessary to deliver a product that truly reflects the population it seeks to serve. Here are a few steps leaders can take to build a diverse dataset from the ground up:

  1. Challenge assumptions about the data you already have. Even if you already have a product, you can take steps to improve accuracy and quality by conducting an audit of existing datasets for biases and gaps in representation across demographics. The accuracy of a model depends on a feedback loop of results that generate adjustments in data collection methodologies and performance. Conducting this type of audit will help you improve performance and set a benchmark for measuring your progress over time as your impact grows.

  2. Be transparent about data handling, research methodology and product validation. At my company, we developed a highly researched vocal biomarker technology and are using it to impact the mental health of millions. To do this responsibly, we have set a new standard of research and validation that involves cross-demographic analysis of our models, testing their accuracy on novel datasets. We frequently publish the results of our clinical and technical studies which detail our scientific methods and approaches to building our AI solutions.

  3. Build models that can be easily updated over time. The needs you may be trying to meet may shift—for example, you may be focusing on a specific population—or the field may advance as you’re building your solution. If you’re generating insights using an outdated model in a new paradigm, you could risk doing more harm than good. In healthcare, staying abreast of the shifting landscape and adjusting as necessary is a standard practice. The same principle applies to datasets. Much the same as building in diverse representation from the ground up, you should always seek to adapt to new standards and be dynamic when it comes to your solutions.

Finally, and perhaps most importantly, we must always remember the Hippocratic Oath: “First, do no harm.” This principle has guided those who practice medicine since the time of Hippocrates. As we look into the future that AI promises us, we should honor the foundations upon which this revolution has been built and commit to using this technology responsibly, transparently and with the betterment of the human race as the defining objective.


bottom of page