Are you Really Ready to Use AI When it Comes to Data-Driven Healthcare Research? A Statistician’s Perspective

The following is a guest article by Justin Manjourides, Associate Professor, Biostatistics, Real World Evidence, Health Data Science at Northeastern University

Relying on AI for data-driven insights is remarkably similar to the traditional statistical methods that have been used to extract meaning from data for hundreds of years. When trying to incorporate emerging new technologies like generative AI into your organizational processes, it is important to not lose sight of all we have learned about the scientific process of generating reliable results. From my point of view as a statistician, here are five things to ask yourself when using AI to inform healthcare research:

1. Will AI Answer the Question that You Actually Want to Ask?
2. Do You Have the Appropriate Data to Answer Your Question?
3. Do You Have Enough Data to Fit the Model You Want to Fit?
4. Are You Aware of the Limitations of the AI Algorithm You’ve Selected?
5. Is the Evidence or Insight Being Generated Reliable?

Will AI Answer the Question that You Actually Want to Ask? 

The first step of any good study is asking the right question. Are you asking the question in a way that will give you the answer you want? There’s a difference between asking “Will the demand for semaglutide injectables rise in 2025?”, “What do we project the demand for semaglutide injectables will be in 2025?”, and “What factors will drive the changes in demand for semaglutide injectables, and by how much?” While all three of these questions are clearly related, the approach we take to arriving at a data-driven answer to each of these questions would be quite different, and an “AI” approach may not be appropriate for each.  

Often, by spending the time to ask appropriate and thoughtful research questions, the analytic choices become clear. Without a clear question, we can sometimes find ourselves with a tool (eg Gen AI) that we want to use, and can only hope that it will be appropriate for our purposes. In the end, the results we end up with may not answer the question we actually are concerned with. As the inimitable statistician John Tukey once said, “Far better an approximate answer to the right question, which is often vague, than an exact answer to the wrong question, which can always be made precise.”

Do You Have the Appropriate Data to Answer Your Question? 

All models, from linear regression to generative AI, are only as good as the data we feed them. All introductory statistics classes will caution you to never extrapolate a linear model beyond the limits of your data set (e.g. your results estimating the relationship between height and weight might not hold for a theoretical 9-foot-tall person). This issue becomes more crucial as we move towards the more “black box” type models, like generative AI and neural networks. With less ability to see how the model is using data, we must ensure that our data are appropriate to answer the question we are asking. By “appropriate,” I mean are they accurate, well-labeled, and do they contain the diversity to fully capture your study population?

Diversity of the data is often the most important aspect of an analysis that can’t be fixed after the fact. We have many statistical methods to handle mismeasured or missing data, but if the data we are using are not representative of the population we wish to describe, more data will just give us a more precise, but incorrect answer. 

Do You Have Enough Data to Fit the Model You Want to Fit?

Many modern AI models require significant amounts of data to train. But it’s not as simple as knowing the exact number of subjects, records, or images that will be needed to adequately train a model. As a statistician, I can give two pieces of good news though: 1) A more diverse dataset is effectively larger in “size” than a dataset where most of your data points look the same. 2) If you don’t have enough data (yet) to fit the fancy AI model of the moment, there are plenty of well-validated, reliable, and trustworthy statistical models that can work for both extremely large and extremely small sample sizes. Nonparametric models, for example, rely on very few assumptions about your data, can be very powerful, and have been used consistently since the early 1700s. 

Are You Aware of the Limitations of the AI Algorithm You’ve Selected?

If you’re asking a “Why” question, will your algorithm give you this type of answer? Often, we get completely focused on “prediction” as a measure of success for models. While accuracy in predictions may be appropriately prioritized at all costs for certain applications (eg self-driving cars), understanding, visualizing, or quantifying the relationships within your data may be the more impactful insight you are after. 

Understanding the limitations of a specific model can help inform the ways in which we design our studies, pre-process our data, and communicate the results of an analysis. As George Box said, “All models are wrong, but some are useful.” Understanding why your particular model is wrong will give context to the results that are generated and can be especially helpful in explaining any surprising results.

Is the Evidence or Insight Being Generated Reliable? 

What makes reliable evidence, especially when we know we’re never going to get the 100% correct answer all the time? Most statistical models always include measures of the uncertainty in the generated results, either in the forms of confidence intervals, prediction intervals, alpha levels, and standard errors. These metrics can help provide context into the degree to which we can act on our results. As data modeling becomes more complex, and given that many of the tools that can efficiently mine enormous datasets for insight are black boxes, how confident can we be with the results? Understanding how your AI models perform across different datasets, in the hands of different users, and with subtle changes to tuning parameters and starting conditions can begin to provide a picture of the reliability, robustness, and reproducibility of your model. Almost any data plugged into almost any model will produce a result, being able to measure and describe measures of reliability is the first step towards generating trustworthy evidence at scale. 

About Justin Manjourides

Justin Manjourides is a biostatistician and observational health researcher helping cultivate the next generation of real-world evidence (RWE) researchers. He is a co-PI for the PROTECT, CRECE, and ECHO Centers at Northeastern. He holds a PhD in biostatistics and an AM from Harvard University and received bachelor’s degrees in mathematics and statistics from the University of Florida.

The following is a guest article by Justin Manjourides, Associate Professor, Biostatistics, Real World Evidence, Health Data Science at Northeastern University

Relying on AI for data-driven insights is remarkably similar to the traditional statistical methods that have been used to extract meaning from data for hundreds of years. When trying to incorporate emerging new technologies like generative AI into your organizational processes, it is important to not lose sight of all we have learned about the scientific process of generating reliable results. From my point of view as a statistician, here are five things to ask yourself when using AI to inform healthcare research:

1. Will AI Answer the Question that You Actually Want to Ask?
2. Do You Have the Appropriate Data to Answer Your Question?
3. Do You Have Enough Data to Fit the Model You Want to Fit?
4. Are You Aware of the Limitations of the AI Algorithm You’ve Selected?
5. Is the Evidence or Insight Being Generated Reliable?

Will AI Answer the Question that You Actually Want to Ask? 

The first step of any good study is asking the right question. Are you asking the question in a way that will give you the answer you want? There’s a difference between asking “Will the demand for semaglutide injectables rise in 2025?”, “What do we project the demand for semaglutide injectables will be in 2025?”, and “What factors will drive the changes in demand for semaglutide injectables, and by how much?” While all three of these questions are clearly related, the approach we take to arriving at a data-driven answer to each of these questions would be quite different, and an “AI” approach may not be appropriate for each.  

Often, by spending the time to ask appropriate and thoughtful research questions, the analytic choices become clear. Without a clear question, we can sometimes find ourselves with a tool (eg Gen AI) that we want to use, and can only hope that it will be appropriate for our purposes. In the end, the results we end up with may not answer the question we actually are concerned with. As the inimitable statistician John Tukey once said, “Far better an approximate answer to the right question, which is often vague, than an exact answer to the wrong question, which can always be made precise.”

Do You Have the Appropriate Data to Answer Your Question? 

All models, from linear regression to generative AI, are only as good as the data we feed them. All introductory statistics classes will caution you to never extrapolate a linear model beyond the limits of your data set (e.g. your results estimating the relationship between height and weight might not hold for a theoretical 9-foot-tall person). This issue becomes more crucial as we move towards the more “black box” type models, like generative AI and neural networks. With less ability to see how the model is using data, we must ensure that our data are appropriate to answer the question we are asking. By “appropriate,” I mean are they accurate, well-labeled, and do they contain the diversity to fully capture your study population?

Diversity of the data is often the most important aspect of an analysis that can’t be fixed after the fact. We have many statistical methods to handle mismeasured or missing data, but if the data we are using are not representative of the population we wish to describe, more data will just give us a more precise, but incorrect answer. 

Do You Have Enough Data to Fit the Model You Want to Fit?

Many modern AI models require significant amounts of data to train. But it’s not as simple as knowing the exact number of subjects, records, or images that will be needed to adequately train a model. As a statistician, I can give two pieces of good news though: 1) A more diverse dataset is effectively larger in “size” than a dataset where most of your data points look the same. 2) If you don’t have enough data (yet) to fit the fancy AI model of the moment, there are plenty of well-validated, reliable, and trustworthy statistical models that can work for both extremely large and extremely small sample sizes. Nonparametric models, for example, rely on very few assumptions about your data, can be very powerful, and have been used consistently since the early 1700s. 

Are You Aware of the Limitations of the AI Algorithm You’ve Selected?

If you’re asking a “Why” question, will your algorithm give you this type of answer? Often, we get completely focused on “prediction” as a measure of success for models. While accuracy in predictions may be appropriately prioritized at all costs for certain applications (eg self-driving cars), understanding, visualizing, or quantifying the relationships within your data may be the more impactful insight you are after. 

Understanding the limitations of a specific model can help inform the ways in which we design our studies, pre-process our data, and communicate the results of an analysis. As George Box said, “All models are wrong, but some are useful.” Understanding why your particular model is wrong will give context to the results that are generated and can be especially helpful in explaining any surprising results.

Is the Evidence or Insight Being Generated Reliable? 

What makes reliable evidence, especially when we know we’re never going to get the 100% correct answer all the time? Most statistical models always include measures of the uncertainty in the generated results, either in the forms of confidence intervals, prediction intervals, alpha levels, and standard errors. These metrics can help provide context into the degree to which we can act on our results. As data modeling becomes more complex, and given that many of the tools that can efficiently mine enormous datasets for insight are black boxes, how confident can we be with the results? Understanding how your AI models perform across different datasets, in the hands of different users, and with subtle changes to tuning parameters and starting conditions can begin to provide a picture of the reliability, robustness, and reproducibility of your model. Almost any data plugged into almost any model will produce a result, being able to measure and describe measures of reliability is the first step towards generating trustworthy evidence at scale. 

About Justin Manjourides

Justin Manjourides is a biostatistician and observational health researcher helping cultivate the next generation of real-world evidence (RWE) researchers. He is a co-PI for the PROTECT, CRECE, and ECHO Centers at Northeastern. He holds a PhD in biostatistics and an AM from Harvard University and received bachelor’s degrees in mathematics and statistics from the University of Florida.