Using Large Datasets to Improve Health Outcomes

Guest post by Lyn Hardy, PhD, RN, Program Officer, Division of Extramural Programs, National Library of Medicine, National Institutes of Health.

Before the advent of algorithms to determine the best way to treat and prevent heart disease, a health care provider looking for best practices for their patients may not have had the resources to find that best method. Today, health care decision-making for individuals and their health care providers is made easier by predictive and preventive models, which were developed with the goal of guiding the decision-making process. One example is the Patient Level Prediction of Clinical Outcomes and Cost-Effectiveness project led by Columbia University Health Sciences.

These models are created using computer algorithms (a set of rules for problem-solving) based on data science methods that analyze large amounts of data. While computers can analyze facts within the data, they rely on human programming to define what pieces of data or what data types are important to include in the analysis to create a valid algorithm and model. The results are translated into information that health care providers can use to understand patterns and provide methods for predicting and preventing illness. If a health care provider is looking for ways to prevent heart disease, an accurate model might describe methods—like exercise, diet, and mindfulness practices—that can achieve that goal.

Algorithms and models have benefited the world by using special data science methods and techniques to understand patterns that guide clinical decisions, but identifying data used in their development still requires practitioners to be conscious of the results. Research has shown that algorithms and models can be misleading or biased if they do not account for population differences like gender, race, and age. These biases, also known as algorithmic fairness, can adversely affect the health of underserved populations by not giving individuals and health care providers information specific to and that directly addresses their diversity. An example of potential algorithm bias is creating an algorithm to treat hypertension without including variated treatments for women or considering life-related stress or the environment.

Researchers are focusing on methods to create fair and equitable algorithms and models to provide all populations with the best and most appropriate health care decisions. Researchers in our NLM Extramural Programs analyze this data through NLM funding opportunities that foster scientific inquiry so we better understand algorithmic effects on minority and marginalized populations. Some of those funding opportunities include NLM Research Grants in Biomedical Informatics and Data Science (R01 Clinical Trial Optional) and the NIH Research Project Grant (Parent R01 Clinical Trial Not Allowed).

NLM is interested in state-of-the-art methods and approaches to address problems using large health data sets and tools to analyze them. Specific areas of interest include:

• Developing and testing computational or statistical approaches to apply to large or merged health data sets containing human and non-human data, with a focus on understanding and characterizing the gaps, errors, biases, and other limitations in the data or inferences based on the data.
• Exploring approaches to correct these biases or compensate for missing data, including introducing debiasing techniques and policies or using synthetic data.
• Testing new statistical algorithms or other computational approaches to strengthen research designs using specific types of biomedical and social/behavioral data.
• Generating metadata that adequately characterizes the data, including its provenance, intended use, and processes by which it was collected and verified.
• Improving approaches for integrating, mining, and analyzing health data in a way that preserves that data’s confidentiality, accuracy, completeness, and overall security.

These funding opportunities encourage inquiry into algorithmic fairness to improve health care for all individuals, especially those who are underserved. By using new research models that account for diverse populations, we will be able to provide data that will support the best treatment outcomes for everyone.