In collaboration with the Amsterdam Data Science Center, we will combine advanced epidemiological and statistical methodology with novel artificial intelligence solutions. This will lead to new methodology to solve a widely recognized problem in epidemiology and data science; and provides the data scientist with a network and methods from outside the data science field. More attention to the reliability of data and causal relationships within datasets will be crucial for future models predicting all sorts of medical conditions in and outside clinical settings, and both for medical statistics and for artificial intelligence.

Our datasets consist of the cohorts hosted by the Amsterdam Public Health and Amsterdam Cardiovascular Sciences research institutes of the Amsterdam UMC. The data of these cohorts are directly linked to real patients and citizens, as well as various registries, with high relevance for health care, prevention and medical development. A list of these cohorts may be found here: https://www.amsterdamumc.org/en/research/institutes/amsterdam-public-health/strengths/aph-cohorts.htm

Envisioned projects

Accurate measurement is key in all research. Measurement of the characteristics of study participants is important for diagnostic and prognostic research, and to predict the occurrence of diseases in the near and distant future. Accurate measurement is also important for  the definition of outcomes and variables in epidemiological studies, and to reliably establish causal relationships.

Causal inference

Among the many machine learning methods proposed to discover strong but unexpected heterogeneous treatment effects, causal forests stand out for their provably reliable confidence intervals. In the first subproject, we will investigate whether they are suitable for detecting treatment effect heterogeneity in cohort data. Because the APH/ACS cohort data is rich in covariates (features), a random forest based solution is very appealing. Particular attention will be paid to their sensitivity to unmeasured confounding and ability to incorporate longitudinal data.

Selection bias and covariate shift

If the data to train a model on are a skewed representation of the target population or target data, then the model may suffer from selection bias. For example, when the admission to the cohort or dataset depends on the availability of model predictors. A similar problem arises when participants become missing over the course of the follow-up period, for example due to mortality or study-fatigue. When the training-set and the test-set for developing a model originate from the same dataset, covariate shift may occur, which may lead to a biased model as well. This problem is extremely relevant for longitudinal cohorts, especially when multiple measurements are being used. In one of the envisioned projects, the new staff member will investigate selection bias in several Amsterdam UMC cohorts and develop AI-driven solutions for it.

Record linkage as a solution

Besides solutions that may be intrinsically linked to the data-science methodology, a solution may also be sought outside the dataset at hand, for example by linking with other data. In our cohort studies, genetic data may be linked to medical data; or data from cancer registries or Statistics Netherlands. In the third sub-project, the researcher will collaborate with our ‘Krachtige Koppelingen’ project and compare different solutions – both from traditional statistics and from AI – to link different types of data to come to causal inferences.

The results of these projects will be used as starting point for subsequent funding requests.