Work packages

WP 1: AI for medical images: Automated screening of diabetic retinopathy

WP-leaders: Benjamin Ricaud and Michael Kampffmeyer, Machine Learning Group, NT-Fak, UiT

Clinical supervisor: Geir Bertelsen, Senior doctor, UNN

This work package is in close collaboration with the project SFI Visual Intelligence, based at the Machine Learning Group at NT-Fak, UiT.

The main task for WP1 is to develop algorithms for identifying diabetic retinopathy. The digital database from retinal images comes from Tromsø Eye Study, which is led by Geir Bertelsen. The access to a longitudinal and representative population-based database of medical images is quite unique and gives this WP a great advantage in the development of algorithms with high internal and external validity.

Geir Bertelsen is also senior doctor at the Department of Ophtalmology, UNN and is currently working with his PhD at the Department of Community Medicine, UiT.

WP 2: Computational Pathology

WP leaders

Lars Ailo Bongo - professor Department of Computer Science

Thomas Karsten Kilvaer - associate professor Department of Clinical Medicine, consultant Oncologist, Department of Oncology, UNN

Lill-Tove Rasmussen Busund - professor Department of Medical Biology, consultant Pathologist Department of Clinical Pathology, UNN

Positions: PhD student starting 01.02.2022.
Clinical pathologist (20%). Position has been announced.

Work package 2 will develop new AI solutions for digital pathology. Our aim is to provide prognostic and predictive biomarkers for clinical pathology. We are especially interested in providing clinically relevant information about the immune cells and their interaction with the tumor. To address this challenge, we use an interdisciplinary approach that combines computer science, machine learning, and pathology. We develop new algorithms and deploy these as interactive graphical tools that can be evaluated by pathologists. We work with hematoxylin & eosin whole-slide images, which are the most commonly used slides in clinical pathology. The outcomes are novel methods for personalized medicine that enable faster and more precise diagnostics. In the long term, we aim to integrate computational pathology methods with other data sources such as clinical findings, radiography images, lab results, and omics data.

WP3: AI-derived clinical decision support based on registry data

I denne arbeidspakken er det planlagt to PhD prosjekter:

1) Utvikling av beslutningsstøtteverktøy for ryggpasienter (data fra Det norske ryggregistret og DIPS)

Prosjektleder: Professor Tor Ingebrigtsen

Biveileder: Professor Tore Solberg

2) Utvikling av beslutningsstøtteverktøy for pasienter med anal inkontinens (data fra Register for anal Inkontinens og DIPS)

Prosjektleder: Professor Stig Norderval

Biveileder: ikke avklart

I begge prosjektene er det behov for nært samarbeid med DIPS om automatisk oppdatering av registrene fra pasientjournalsystemet (DIPS).

Det har vært utlyst stillinger for LIS leger i kombinert utdanningsstilling og stipendiatstilling (dobbeltkompetansestilling) for begge prosjektene. 50% stipendiat og 50% LIS-lege utdanningsstilling skal føre fram til dr.grad og klinisk spesialitet innen 6-8 år.

Anal inkontinensprosjektet har kandidat til stillingen.

WP4: AI-derived prediction of adverse events

Many high-risk operations result in some form of adverse event. Early warning of adverse events can minimize the number of complications and lead to fewer readmissions.

Wp leaders

Karl Øyvind Mikalsen, associate professor, Dept. Clinical Medicine and ML Group, Centre manager - Centre for Clinical Artificial Intelligence (SPKI), UNN, www.spki.no

Cristina Soguero, Associate professor, Rey Juan Carlos University, Madrid, Spain

The overarching idea of this work package is that early warning of postoperative adverse events may be realized based on preoperative, perioperative, and postoperative data from heterogeneous sources, such as for example free text from clinical notes, blood samples over time, and other types of structured data such as diagnosis and procedure codes.

The theoretical approach to develop reliable models for prediction and prevention of postoperative adverse events consists of three tasks that address concrete methodological knowledge gaps, namely

to develop new transparent, interpretable, and explainable AI models for EHR predictive models;
to develop new unsupervised and weakly supervised ML methodology in order to leverage vast amounts of unannotated or noisy labeled EHR data;
to develop new ML methodology capable of exploiting contextual and prior information by harnessing the unique hierarchical nature of EHR data sources.

This research is planned to lead up to an online decision support system for daily clinical use.

Positions: Researcher starting in January 2022.

WP 5: Comparative and action-based research on implementation of AI into clinical practice

The work package is led by professor Turid Moldenæs and assistant professor Hilde Marie Pettersen, and is a part of the research group in organization and leadership at the Faculty of Humanities, Social Sciences and Education, Department of Social Sciences https://uit.no/research/fiol.The work package includes a Ph.D. position, which starts 01.01.2022.

There are high expectations to the impact and socio-economic benefits of AI for healthcare professionals and patients. Knowledge on successful implementation is key to unlocking the potential of the technology, for which there is a need to identify barriers and enablers in the implementation process and at structural levels in the hospital, e.g., strategic, operational and frontline.

The overarching goal of the pillar is to identify inhibitors and enablers in the implementation process at different technological and structural levels in the hospital and use this knowledge to facilitate the successful implementation of AI technology in clinical practice.

Technology adoption and implementation involves a complex interplay between organizational change, redesign of work processes and the technology itself. From the literature about how implementation is organized we know that the key variables for outcomes are duration and pressure of attention, training, follow-up and control. This literature further provides reasons for assuming that at least three conditions are of importance to understand how physicians who are supposed to adopt the new technologies will respond. The first concerns physicians` clinical judgment, which could be challenged by technology, and thus be a barrier to implementation. The second is the traditional evidence orientation so strongly rooted in medicine, which can be an enabler to implementation. However, it requires that physicians are convinced through scientific studies that the technology works. It is also possible that AI and clinical judgment will be mixed and used simultaneously. Finally, leaders' competencies, attitudes, intentions and behaviors influence technology adoption.

The literature on readiness for change, on the other hand, emphasizes how much organizational members value the change and how favorably they appraise three key determinants of implementation capability: task demands, resource availability, and situational factors. This literature also makes a distinction between barriers and enablers on different organizational levels; strategic, operational and frontline level. Accordingly, AI as other new technologies, cannot be expected to be implemented into routine clinical practice unless the organizational context is taken into consideration, and actively managed by leaders at all organizational levels.

By combining the two aforementioned research approaches we; a) aim to identify organizational requirements for successful implementation of AI in the clinical routine work, and b) use this knowledge to strengthen the implementation of the developed AI-systems in the project.

The study will be based on a comparative research case design (departments/working groups that are implementing the AI-technology and departments/working groups working more traditionally).