CORDIS - EU research results

Personalised Medicine by Predictive Modeling in Stroke for better Quality of Life

Periodic Reporting for period 3 - PRECISE4Q (Personalised Medicine by Predictive Modeling in Stroke for better Quality of Life)

Reporting period: 2021-05-01 to 2022-10-31

Stroke is one of the most severe medical problems with far-reaching public health impact. PRECISE4Q was set out to minimise the burden of stroke for the individual and for society. It has created multi-dimensional data-driven predictive computer models enabling personalised stroke treatment, addressing patient’s needs in four stages: prevention, acute treatment, rehabilitation and reintegration.

* Heterogeneous data from multidisciplinary sources has been integrated: genomics, microbiomics, biochemical; imaging including mechanistic biophysiological models of brain perfusion/function; social, lifestyle, gender; economic and worklife, requiring substantial efforts for information extraction, semantic labelling and standardisation.
* The predictive capability and clinical precision has been validated with real clinical data generated by (i) clinical studies and (ii) retrospective analyses of big data sets.
* In a value-based pricing system, PRECISE4Q, could generate massive reductions in direct and indirect health care costs as personalised prevention, treatment, rehabilitation programmes and reintegration leads to fewer cases of stroke, better procedures, better outcomes for patients and higher return to work rates, respectively.
Institute Guttmann and Technological University Dublin used multi-task learning (MTL) and the Barthel Index to predict stroke patient Independence after rehabilitation. The AI models predict the components of the BI and demonstrate benefits of MTL in a health context to predict patient profiles. Given the proliferation of multi-component assessment tools across medicine, the use of multi-task learning to improve predictions of patient outcomes has the potential to have a significant impact across many areas of health, thereby adding value in all of these areas for stakeholders, and society. The models will be used for validation in clinical settings in stroke rehabilitation and reintegation. Several scientific papers has been published in the models.
Together with University College Dublin, the both beneficiaries also developed a predictive model for individual reintegration and well-being profiles after rehabilitation.
Institute Guttmann also advanced their "Guttmann Neuropersonal Trainer: GNPT is a web-based platform including two main modules: an interface for healthcare professionals providing management tools and a module for patients including a set of cognitive tasks addressing the main cognitive functions involved in activities of daily living. Guttmann designed, implemented and validated predictive models to be integrated in GNPT addressing two main clinical needs: cognitive improvement and therapy compliance. Such models were required by clinicians to support them in optimizing their current cognitive therapy targeting the main cognitive functions involved in daily living activities (e.g. memory, attention or executive functions).
Furthermore, Institute Guttmann developed personalised rehabilitation models for post-stroke patients. Predictive models were implemented for the personalised stroke rehabilitation stage within the PRECISE4Q project. The three dedicated models are i) the cognitive improvement model, ii) the cognitive therapy compliance model, and iii) the motor improvement model.

Medical University Graz developed multilingual interface terminologies to link terminology standards of health records such as SNOMED CT, with clinical jargon expressions as found in clinical documents. As a result, several million of German language clinical terms were linked to SNOMED CT, and can be used for entity normalization in clinical text analytics.
Medical University Graz also developed an innovative approach to medical data harmonization and integration. The approach is agnostic to any specic healthcare data standard representation but bears the potential to bridge across heterogeneous standardized and non-standardized representations of clinical data. It is supported by a semantic repository, providing a harmonized view on data that are heterogeneous regarding scope, granularity, provenance, and representation standard.

QMENTA integrated structured and unstructured data from multiple countries in their imaging and stroke data workflow platform. The platform can be used to develop and run analysis on imaging data.

DFKI developed an AI based multi-lingual text mining model for digital health data. Deep Learning methods were developed for: Knowledge Graph Completion and Embeddings; Named Entity framework based on transfer learning (called T2NER); MedDistant19; MEDDISTANT19; a benchmark for broad-coverage biomedical relation extraction; and a method for few-shot cross-lingual transfer for coarse-grained de-identification of code-mixed clinical texts.

Charité advanced a tablet-based clinical decision support for optimizing treatment decisions by doctors treating acute stroke patients (called 'STAR'). STAR is based on cutting-edge AI (artificial intelligence) models trained on large clinical datasets. In contrast to current generalized standard of care, our solution provides a personalized treatment recommendation for best individual and measurable outcome in terms of survival and quality of live. The underlying models integrate clinical, epidemiological and imaging data of the individual patient such as history, co-morbidities or medication.
Linköping University developed a Digital Twin prototype for new eHealth solution. This Digital Twin can simulate what happens in the human body in response to basic interventions: exercise, change in diet, and certain medications. The world-unique aspect of our simulations is that they describe the interplay between detailed organ models, for almost all of the main organs in the human body. These models are transferred to a backend platform. This backend communicates with two frontend prototypes: one for in-depth simulations in connection to e.g. teaching, or conversations with a doctor or teacher, and one for usage in a home environment, to help keep goals regarding training, diets.
Finally, University of Tartu developed individualised risk prediction models for individuals at risk of stroke. Genetic, life-style and health related data from the Genome wide association study meta-analysis for stroke were used to build polygenic risk scores (iPGS) for stroke and its riskfactors. These were tested in the Estonian Biobank and cohorts with other ancestries.
A total of 38 scientific papers were published based on the Precise4Q project results with an additional 13 publications which are related to the project but have not acknowledged funding from Precise4Q. All publications can be found on the project website.
PRECISE4Q has provided true personalized predictive modelling in stroke across four stroke phases: prevention, acute stroke, rehabilitation and reintegration. To this end, we have innovated by developing (1) novel hybrid model architectures, (2) novel complex models (including deep-learning and gradient boosting models) capable of (3) feature ranking to inform novel intervention design, (4) scaling to high-dimensional data and big-data and integrating different forms of input data (both image and structured), and (5) predicting complex interrelated and consistent structures using multi-task learning and structure prediction methods.