Back to search

IKTPLUSS-IKT og digital innovasjon

Deep Learning for Health

Alternative title: Dyp læring for helse

Awarded: NOK 0.59 mill.

This project will advance the research field of deep learning within the context of data analysis in healthcare using data from electronic health records (EHRs) and medical imagery. EHRs store digitalized data about patients, including free text notes, blood sample results, diagnosis codes and procedure codes. The EHR source of data is as of today largely untapped when it comes to researching predictive models for risk assessment, or models for diagnosis and decision support. The project will revolve around uniquely available data from EHRs related to gastrosurgery as well as medical imagery, with a special emphasis on colorectal cancer. The aim is to develop entirely new analysis tools for prediction of postoperative complications. Such complications, including for instance wound infections and anastomosis leakage, occur in around 25% of patients undergoing high-risk operations, increasing the chance of death within 5 years. Prediction of postoperative complications may reduce the rate of complications by early intervention, leading to better quality of life for patients, better quality of surgery, as well as reduced costs within the healthcare system. To be able to extract as much information as possible for predicting postoperative complications requires entirely new analysis tools. Towards that end, the project will embark upon foundational research to advance deep learning methodologies. Deep learning is based on training enormous artificial neural networks from data to perform predictions, and has provided a revolution in fields such as speech recognition and image analysis. Similar breakthrough performance is realistic within health, but requires basic research in order to make deep learning more robust with respect to the size of the data sets under study, to better handle missing data, and to better fuse information from different data sources, while at the same time providing explainable models. The use of the data is approved by REK.

DeepHealth Phase 1 made progress towards developing entirely new ways to analyse patient data from electronic health records. This project will disseminate some of these results (concretely, present four peer reviewed papers at international conferences), as well as further develop some of these results/prototypes with the aim to investigate opportunities for building on these works for new funding proposals (e.g. H2020). Background: Vast amounts of heterogeneous and complex data from Electronic Health Records (EHRs) are ubiquitously being recorded at the patient level in healthcare (big data). This represents a largely untapped source of data-driven clinical information, having the potential to transform health and leap forward quality of care for the individual patient. However, this requires inference tools of much greater sophistication than traditional tools that often suffer from weaknesses such as oversimplified modeling and predictions based on population averages. As a future and emerging technology in artificial intelligence and cognitive systems, deep learning has revolutionized analysis of big data in applied domains such as speech and image analysis. Similar breakthrough performance is realistic in health, provided that main challenges related to deep learning, especially in health, are resolved. This included the high dimension-small sample size problem (d>>N), heterogeneous source integration, and missing data. DeepHealth's aim is to move the research front in deep learning and artificial intelligence for data analysis beyond the current state-of-the art for the best quality of care. This will be achieved by a long-term research endeavor within the context of ubiquitous data and services in healthcare for prediction and prevention of postoperative complications, an enormous problem in health. The project will leverage vast amounts of uniquely available EHR data and clinical imagery from the University Hospital of North Norway, related to gastrosurgery and especially colorectal cancer, for which surgery is the only curative treatment. DeepHealth will perform analysis before, under, and after surgery jointly on unstructured and structured data, times series data, and imagery, for predictions of postoperative complications. Close collaboration exists with surgeons and international deep learning and computational health expertise, wherein high mobility will be key.

Publications from Cristin

No publications found

No publications found

No publications found

No publications found

Funding scheme:

IKTPLUSS-IKT og digital innovasjon