Is artificial intelligence a friend or foe to epidemiology?

Abstract

Epidemiology has long been central to public health, guiding our understanding of the distribution and determinants of disease. As the field has evolved—from John Snow’s cholera investigations to large-scale cohort studies and causal inference frameworks—it now faces a transformative juncture with the advent of artificial intelligence/machine learning (AI/ML). These technologies offer unprecedented opportunities to improve data measurement, inference, and population health insights, yet also pose methodological and ethical challenges. Anchored by the core epidemiologic domains of study population, measurement, and inference, we examine how epidemiologists can use AI/ML effectively. We consider the importance of careful population definition, informed sampling, and external validation to ensure generalizability and minimize bias when AI/ML is used. We also explore the need for rigorous assessment of data quality and model reliability, which strengthens the case for conceptual frameworks in guiding interpretation of scientific investigations. To realize AI/ML’s potential, epidemiology must adapt its training, invest in infrastructure, and promote interdisciplinary collaboration. Doing so will ensure that epidemiologic science remains robust, reproducible, and relevant in a rapidly evolving informational landscape. This moment calls for a strategic integration of AI/ML into the fabric of epidemiologic practice and training to advance both science and public health.

Background

For centuries, epidemiology has served as a cornerstone of public health prevention and practice, providing critical insights into the distribution, patterns, and determinants of health and disease. The field has progressed over time, with key inflection points marking changes in epidemiologic theory, approaches, and tools. John Snow’s work on cholera in 1854 challenged miasma theory and paved the way for germ theory and the development of epidemiologic principles still used today [1]. W.E.B. Du Bois’ seminal work, The Philadelphia Negro, employed mixed methods to study the health effects of racism and other social factors as early as 1899 [2]. In the mid-20th century, large-scale cohort studies like the Framingham Heart Study, the British Doctors Study, and the Seven Countries Study built our knowledge base of risk factors for chronic disease [3]. Landmark tobacco studies have exemplified the power of case-control investigations and principles of causal inference [4]. Since then, epidemiology has continued to evolve while accommodating broad new areas of inquiry such as those focused on examining the effects of genetics and social determinants of health on the well-being of populations.

Epidemiology now stands at an important crossroads as the field confronts the rapid expansion of big data and artificial intelligence/machine learning (AI/ML) [5]. The availability of large, heterogeneous datasets has accelerated the development of ML, deep learning (DL), and large language models, enabling advances in tasks such as clinical text summarization, diagnosis, surveillance, and genomic analysis [6–9]. For example, infectious disease experts have used AI/ML to rapidly query surveillance data during outbreaks [10]. Cancer epidemiologists have used DL methods to expand existing Cox regression models by leveraging advantages of neural networks [11]. Pharmacoepidemiologists have deployed natural language processing (NLP) techniques, ranging from rules-based models to large language models, to improve exposure and outcome ascertainment within electronic health record data [12]. Emerging agentic AI systems extend these capabilities beyond passive prediction toward limited forms of planning and task execution [13]. Together, these developments expand and accelerate analytic possibilities but simultaneously raise new methodological and ethical challenges for epidemiology.

The discipline now faces key questions as to whether, and how, these new technologies can best be integrated and embraced. Will epidemiologists modify their tools and methods to incorporate AI/ML into research, training, and service, or will these advances remain largely in the realm of informaticians, computer scientists, and others adjacent to the core practice of the profession?

We ground our considerations in the tenets of epidemiology: study population, measurement, and inference (Fig. 1). We then discuss steps the field must take to ensure that the value of AI/ML, as tools and drivers of inquiry, are fully realized as the practice of epidemiology continues to advance.

Fig. 1.

Validity considerations in epidemiologic studies using AI/ML. Adapted from 340.721.81 Epidemiologic Inference in Public Health I. Lecture on Populations. Celentano (2024).[55]

Key definitions

To support a common vocabulary for an epidemiology audience, we define the following terms used throughout this commentary (Table 1). The majority of the AI approaches delineated here are developed using supervised, unsupervised, or reinforcement learning paradigms trained on a myriad of data types [17].

Table 1.

Key definitions

Term	Definition
Artificial intelligence	Technologies and systems capable of performing tasks traditionally done by humans, such as pattern identification, image classification, or information summarization [14].
Machine learning	A core subfield of AI that comprises approaches that allow machines to learn from data and subsequently make accurate inferences about new data [15].
Deep learning	A branch of ML that uses layered artificial neural networks to iteratively process and transform input data through mathematical operations [16].
Natural language processing	AI methods used to interpret, structure, and generate human language.

Abbreviations: AI, artificial intelligence; ML, machine learning

Study population

To successfully answer an epidemiologic question, one must identify the target, source, and study populations. This process undergirds external validity, allowing inferences from a study population to the target population as well as application of findings to a new target population. Such efforts are equally if not more important in the AI/ML era. Carefully defining populations in terms of person, place, and time is particularly relevant when developing predictive models, classification systems, and other ML algorithms that may use a large corpus of data from multiple sources. Data diversity in ML captures the breadth of data elements spanning multiple types, sources, and contexts. However, this concept does not automatically translate to demographic or clinical diversity of individuals represented in data, and an uninformed sampling strategy can obscure intended target populations. Epidemiologists are equipped to explore and synthesize diverse data, drawing inferences from rows of observations while carefully acknowledging who the observations do and do not represent.

Once a target population is defined, one can then evaluate whether the available data represents or generalizes to all members. Informative sampling, or sampling with a clear understanding of the target population and inferential goals, is necessary to train and evaluate AI/ML models appropriately. For example, performance measures such as loss functions can amplify concerns of misrepresentation if interpreted or used incorrectly. Most standard loss functions like mean squared error or log loss are global averages across individuals or observations, meaning that groups with more data contribute more to the loss [18]. As a result, typical AI/ML models optimize their performance to the majority groups, potentially at the expense of minority or underrepresented groups, compromising generalizability. Epidemiologists can help determine when results are generalizable through informative sampling or weighting to align key covariate distributions between the study and target populations. This process requires accurate measurement of the target population’s characteristics, often informed by population-based surveys or national data systems like a census, or, in cases where data are unavailable, determining whether the target population requires redefinition. Epidemiologists can also assess whether results generalize in interpretation, drawing on knowledge of the same mechanism or biological process in both the study and target populations [19]. Informed sampling and data selection may still present subtle challenges. Arora et al. provide the hypothetical example of a dataset in which members of an ethnic group are numerically representative according to Census data proportions, but whose outcomes are systematically misclassified relative to the sampled population [20], illustrating that even proportionate representation does not preclude misrepresentation.

AI/ML applications must also be tested in different populations, settings and time periods to assess transportability to new target populations. Both discrimination and calibration are important, in addition to standard measures of ML accuracy and precision such as F1 scores. Consider a novel machine learning algorithm to predict waist circumference developed using national survey and trial datasets [21]. If a new study prompted waist circumference evaluation among South Asians, who are not well-captured in these datasets [22] and have differing thresholds for obesity [23], it would be necessary to modify the algorithm for this new target population. In this case, failing to consider the differences between the study population and a new target population could lead to inappropriate use of an algorithm and interpretation of its performance.

To ensure study populations are accurately constructed and described, epidemiologists must critically appraise who is reflected in the data and, perhaps more importantly, who is excluded. Structural barriers contribute to the underrepresentation of specific populations of interest. Specifically, limited healthcare access and legal or ethical restrictions on data sharing and digitization [20] affect the scope and validity of AI/ML applications in epidemiologic research. When applied without careful attention to these limitations, AI/ML methods risk reinforcing or exacerbating existing inequities, highlighting the ethical importance of fairness, privacy, and data protection in algorithm design and implementation [24].

Measurement

Measurement in epidemiologic research has historically been constrained by limited data availability, underpowered sample sizes, and challenges in integrating diverse datasets. AI/ML offers immense potential to address these limitations by harnessing the ever-growing quantities of data to enhance epidemiologic research [25–27]. In 2003, the total amount of data created globally was about 5 exabytes; by 2024, this number skyrocketed to 149 zettabytes, increasing by nearly 30,000-fold [28,29]. While the sheer quantity of data is promising, it does not guarantee quality. It is crucial to understand both the measurement opportunities of AI/ML and the likelihood for measurement error.

If used appropriately, AI/ML can enhance measurement in epidemiology by improving data quality, automating data collection, and refining construct measurements. Advanced imputation techniques have been used to address missingness and identify misclassified variables, inconsistencies, or erroneous values [30]. AI/ML techniques, such as NLP and computer vision, can assist in automating data extraction from diverse sources, including medical records, surveys, or images. This can improve the depth and efficiency of measurement, especially in settings with large volumes of unstructured data.

Before harnessing the power of data obtained by or applied to AI/ML models, epidemiologists must evaluate the validity and reliability. For example, optical heart rate sensors used in AI/ML-enabled wearables often suffer from motion artifacts, which may falsely detect atrial fibrillation and prompt individuals to unnecessarily seek medical attention [31,32]. These sensors collect large volumes of data but can compromise construct and criterion validity. Equally critical, reliability impacts the stability and reproducibility of models, as low measurement consistency introduces noise and contributes to overfitting [33]. Epidemiologists may utilize methods like the Kappa statistic and coefficient of variation to evaluate reliability, alongside AI/ML techniques to examine label agreement and assess test sample loss impact [34]. Both bias and imprecision introduced by measurement error can impact feature selection, commonly used in AI/ML to remove non-informative independent variables. A low signal-to-noise ratio, resulting from either systematic mismeasurement or low precision of measurement, obscures the true effect and may have the undesirable impact of omitting important variables from AI/ML applications. Greater volume and complexity of data does not necessarily increase the validity and reliability, and accurate measurement must be prioritized when using AI/ML tools in epidemiologic work. Better measurement of a health construct may involve combining multiple measurements, multimodal data, and analytical techniques to correct measurement error, particularly when aiming for highly sensitive or reliable measurements.

Finally, it is vital to validate AI/ML tools against a gold standard measurement and in external populations. Validation studies are critical to mitigating information bias by quantifying measurement error relative to a gold standard, which can vary in different populations due to random chance, population characteristics, or study biases. A recent systematic review [35] by Rakers et al. found that approximately 61 % of predictive ML algorithms deployed in primary care lacked information on algorithm external validation, raising concerns about their reproducibility and applicability. Consider IBM Watson for Oncology, a now retired AI tool that provided clinical decision support for cancer diagnosis and treatment [36]. The system was trained on hypothetical cases rather than real data, which contributed to inappropriate treatment recommendations, with concordance varying by country and cancer type [37,38]. Conversely, other early ML tools demonstrate strong internal and external validation. The Framingham risk score, which gave rise to the widely used atherosclerotic cardiovascular disease risk score, was validated against adjudicated cardiovascular events and has since been validated in diverse populations apart from the predominantly white, middle-aged cohort in which it was developed [39–41]. This validation process has built a substantial evidence base supporting the score’s usage and highlighting areas of caution across populations. Newer AI/ML tools must focus on measurement validation, a fundamental epidemiologic principle necessary for inference.

Epidemiologic inference

Epidemiologic inference refers to the development of generalization from sample data across descriptive, predictive, and causal paradigms [42]. Correct inference is rooted in good study design, including sound analytic methods and data integrity, as well as thoughtful interpretation. A well-defined research question and study design are essential for drawing accurate conclusions about covariates and their relationship to an outcome. Specific constructs take on different roles in studies depending on the research question. For example, a construct that is a confounder in one study can be a predictor, mediator, or descriptor in another investigation. Accurate interpretation of findings is contingent upon the study design and its corresponding data structure.

In cross-sectional studies, while data structuring may appear straightforward, framing the results correctly is crucial to acknowledge the study design’s limitations. Otherwise, non-causal explorations may be deemed causal, non-temporal patterns may be perceived as temporal, prevalent cases will be called incident, and so forth. This is especially important when AI/ML is used for hypothesis generation by identifying patterns that may not be evident through traditional data exploration methods. In longitudinal studies, any misspecification of the time-related components can not only increase bias but also completely alter inference. The study design must inform the structure of the analytic data set; for longitudinal studies, key concepts such as time origin, types of censoring and the resulting inferences, and handling of time-varying variables must be reflected in the data structure. Importantly, not all ML algorithms are suitable for modeling the correlated nature, missing measurements, non-linear trajectories, and other complexities of longitudinal data [43]. However, time-to-event models bolstered by ML have increased the flexibility of risk modeling to generate real-world evidence [11]. Many AI/ML programs, code sharing platforms, and analytical tools have streamlined advancements in areas such as computer vision, NLP, and health care diagnostics. However, improper use of the tools may lead to incorrect inference.

Erroneous inference can result from inattention to data structure or the processes that generate data. Insurers commonly use AI-based prediction models for outcomes such as disease onset or hospitalization, and model outputs often guide interventions targeted to “risk factors” or high-risk subgroups [44]. However, predictive features are sometimes misread as causal drivers when they instead reflect bias. Without careful attention to temporality, highly predictive features may reflect markers of incipient disease rather than true antecedent risk factors. These models also implicitly assume complete observability of outcomes in electronic health records or claims [45]. In reality, representation in these sources, and verification of disease, reflect differential access to health care. Furthermore, unequal access to health care by factors like race and socioeconomic status may result in differential missingness or misclassification of disease and clinical indicators used for prediction [44]. Obermeyer and colleagues’ widely cited work demonstrates how algorithms can propagate disparities in care by failing to consider how indicators of health care access, like total costs, may be flawed proxies for actual health status [46].

Epidemiologists can guard against such misinterpretations by providing a map for inference. A conceptual framework guides inquiry and aids interpretability by specifying the hypothesized relationships to be studied. Such frameworks include directed acyclic graphs used in causal inference, as well as other diagrams that visualize etiology, levels of influence, or feedback loops for a given health phenomenon, and make our assumptions explicit. By delineating who is affected, what we measure, and how the phenomena operate, conceptual frameworks connect the concepts of study population, measurement, and epidemiologic inference. AI/ML excels at detecting correlations but cannot by itself provide the contextual explanations that frameworks provide. Despite its challenges, epidemiologists have already embraced AI/ML to enhance inference, with care to address sources of bias at each stage of algorithm development and validation [26,47].

Is the field of epidemiology prepared?

If the field of epidemiology is to continue to fulfill its promise, it is our position that it has no choice but to accommodate the changing world, including the proliferation of tools and methods that support the application of AI/ML. Currently, many epidemiologists and AI/ML experts speak different languages, and there is a general lack of methodological and scientific expertise at the crossroads of the two disciplines. This barrier necessitates clear, plain-language communication between disciplines to establish a common ground. This strategy can reduce misunderstandings in complex scientific collaboration in health research, particularly as large-scale data initiatives proliferate across research universities [48,49].

Epidemiologic training traditionally focuses on descriptive, causal, and implementation investigations [50], whereas AI/ML curricula often emphasize pattern recognition, automation, and prediction with a focus on efficiency and scalability. While epidemiology curricula focus on R, SAS, and Stata for traditional biostatistical analysis, AI/ML largely relies on Python. Knowledge of fundamental programming languages provides one of many critical checkpoints to reaching AI literacy. Many trainees lack programming skills and knowledge of AI terminology [51], which consequently limits their opportunities to gain AI-scientific literacy through advanced coursework, faculty mentorship, and research assistantships. The opaque nature of AI/ML and costs associated with development also increases the natural skepticism among academics, which continues the cycle of epidemiologic isolation from AI innovation [52].

There are three critical steps that departments and schools can take as they invest in new training and collaborations. First, educators must incorporate AI/ML into traditional epidemiology curricula such as courses focused on measurement, study design, and analytic tools for epidemiologic inference. AI training in epidemiologic curricula is imperative for cutting-edge research, while also ensuring ethical and proper use. Fig. 2 presents a sample AI-integrated epidemiology curricula. Investing in the development of AI-trained professionals involves creating training programs that teach ML and Python programming tailored to epidemiologic research. Providing trainees opportunities to work on real-world epidemiologic projects allows them to apply their AI and programming skills to solve complex health problems that can be actualized in theses, capstone projects, and dissertations. These collaborative experiences accelerate up-skilling in key areas like data wrangling, statistical reasoning, reproducible workflows, and responsible AI practice that are required to navigate the rapid changes in the AI landscape. As curricula incorporate AI/ML, educators must emphasize not the tools themselves, but their ability to help answer well-defined and impactful public health questions [53].

Fig. 2.

Example AI curricula for a graduate epidemiology training program. NLP Natural Language Processing; AI Artificial Intelligence.

Second, departments and schools should devote increased resources for trainees and faculty to access big data and AI tools. Institutions should take inventory of resources available to trainees and faculty, including data assets, cluster computing, and experts in AI/ML methodologies. Standard guidance on privacy, ethics, and data governance ensures the responsible use of these resources for education and research.

Finally, leaders should promote interdisciplinary collaboration with software engineers, data scientists, and biostatisticians. Promoting said collaboration can enable epidemiology departments to grow beyond their traditional operating structure. Think tanks focused on AI/ML and epidemiology can bring together experts from epidemiology, computer science, and biostatistics to share insights, inventory, and advancements in AI/ML applications and reduce the barrier of discipline-specific language. These exchanges can happen through organizing workshops, seminars, and conferences. Bringing together diverse expertise fosters sharing AI/ML resources and bolsters reproducibility in the epidemiology community, which in turn paves the way to interpretable and impactful applications of AI.

Conclusion

During the past decade, remarkable advances have been made in AI/ML, ranging from AI-powered medical diagnostics and reinforcement learning to generative AI like ChatGPT [54]. The tenets of epidemiology – study population, measurement, and epidemiologic inference – provide one lens through which to consider the promise, as well as perils, of these advances. Looking ahead, thoughtful integration of AI/ML has the potential to strengthen epidemiologic practice by expanding analytic capacity, improving data workflows, and enabling new forms of collaboration across disciplines. While promising, expectations that these tools will fundamentally redefine how epidemiology generates or validates knowledge have outpaced the evidence base. Before AI/ML is entrusted with reshaping core principles, they must be rigorously evaluated against established methods to understand their benefits, limitations, and trade-offs. However, failing to adapt risks widening disparities in methodological literacy, limiting the field’s ability to evaluate emerging technologies, or ceding interpretive authority to domains less grounded in public health. For departments and schools where epidemiology is taught, these challenges also present opportunities to modernize training, build shared infrastructures, and ensure that epidemiologists remain central to guiding the responsible use of AI/ML.

Our proposed changes to the traditional epidemiologic curricula provide a foundational framework for AI/ML integration; however, programs should continuously update educational objectives and curricula to keep pace with the evolving AI/ML ecosystem. A deliberate and proactive approach will help to maximize the promise of the field in understanding the distribution, patterns, and determinants of health and disease.

Abbreviations:

AI

Artificial intelligence

DL

Deep learning

ML

Machine learning

NLP

Natural language processing