Research
Latest publications in ML/AI
Prediction of Cardiovascular Markers and Diseases Using Retinal Fundus Images and Deep Learning: A Systematic Scoping Review
Li LY, Isaksen AA, Lebiecka-Johansen B, Funck K, Thambawita V, Byberg S, Andersen TH, Norgaard O, Hulman A
European Heart Journal - Digital Health 2024;,ztae068 10.1093/ehjdh/ztae068
Large Language Models for Epidemiological Research via Automated Machine Learning: Case Study Using Data From the British National Child Development Study
Wibaek R, Andersen GS, Dahm CC, Witte DR, Hulman A
JMIR Med Inform 2023;11:e43638 10.2196/43638
ChatGPT- versus human-generated answers to frequently asked questions about diabetes: A Turing test-inspired survey among employees of a Danish diabetes center
Hulman A, Dollerup OL, Mortensen JF, Fenech ME, Norman K, Støvring H, Hansen TK
PLOS ONE 2023;18(8):e0290773. 10.1371/journal.pone.0290773
Perception of artificial intelligence-based solutions in healthcare among people with and without diabetes: A cross-sectional survey from the health in Central Denmark cohort
Schaarup JFR, Aggarwal R, Dalsgaard E-M, Norman K, Dollerup OL, Ashrafian H, Witte DR, Sandbæk A, Hulman A
Diabetes Epidemiology and Management 2023;9:100114, 10.1016/j.deman.2022.100114
Transfer learning for non-image data in clinical research: A scoping review
Ebbehoj A, Thunbo MØ, Andersen OE, Glindtvad MV, Hulman A
PLOS Digit Health 2022;1(2):e0000014. 10.1371/journal.pdig.0000014
Strategic areas
Clinical prediction - using deep learning, where it matters
We are mostly interested in how to integrate multimodal data (e.g. images, time series, text) in clinical risk prediction using deep learning methods in addition to commonly used variables (tabular data). In this process, we see a potential in reusing models developed in other datasets or even domains, to transfer knowledge representations between prediction tasks (e.g. using transfer learning to repurpose generic computer vision models for analysis of retina images).
Knowledge transfer - from arXiv to the clinic
The data science and clinical research communities have very different research cultures. The latest deep learning methods are often introduced at computer science conferences and are described in preprints, which does not reach only a very few clinical researchers. Also, data science methods are often developed using open access datasets (also medical), which don’t necessarily reflect data collected in a clinical setting or the methods don’t address clinically relevant problems. Therefore, there is a need for interdisciplinary groups who have experience in both worlds and can spot both clinical needs and methods to address them. In addition to this, we are planning to conduct scoping reviews about deep learning methods and present them in a language accessible to the clinical research community. Moreover, we will interact with stakeholders (clinician, policy makers) and communicate our findings to patient organisations.
Generative AI - creating resources for further research
The collection of clinical data is often cumbersome and expensive, while data sharing and collaborations are often hindered by challenges with data sharing. Our vision is that generative AI methods (e.g. generative adversarial networks) can be used to create synthetic datasets that mimic real data, but without privacy issues, and sharing such datasets can encourage collaborations and attract the data science community to certain clinical problems in diabetes research.
Experiments - from wild(ish) ideas to quick prototypes
Large language models make the power of AI accessible to a broader user group. We aim to leverage these technologies to quickly build and test prototypes among end users (e.g. in the clinic), before allocating more resources to develop complete and expensive solutions. Such case studies or proof-of-concepts can give us valuable insights about the needs and perception of AI in the clinic.
The impact of AI on healthcare - end users matter
By asking end users (patients, clinicians) about their perception of AI, e.g. hopes and fears, we aim to get valuable insights and inspiration for our work. We plan to do this by conducting large survey studies and arranging user involvement sessions.
Funding
Integration of longitudinal multimodal data in clinical risk prediction using deep learning (2023-2028)
Artificial intelligence enables computer programs to execute human-like tasks like image and speech recognition, text translation, and more. These applications are based on deep learning, a method that can recognize patterns in large datasets (e.g. millions of images from the internet) and then make predictions for new cases. In this project, deep learning methods will be developed and applied in a clinical setting. Persons with type 1 diabetes visit their physicians regularly for check-ups and screening for complications. Some of them also monitor their health using wearable devices even between visits. Combining these data creates a unique opportunity for the development of clinical prediction models that can assist clinicians to tailor prevention and treatment. However, complex data of different types (tabular, images, time series) collected repeatedly over time call for the development and application of novel deep learning methods.
This project is funded by the Novo Nordisk Foundation under the Data Science Emerging Investigator Programme (NNF22OC0076725).
Prenatal immune-modulatory diets and biomarkers in relation to risk of developing type 1 diabetes during childhood (2023-2026)
This project is funded by the Novo Nordisk Foundation under the Steno Collaborative Grant Programme (PI: Flemming Poicot, Steno Diabetes Center Copenhagen).