Work package 3: Algorithm development

Lead: Vrije Universiteit Amsterdam

Objectives

The main objective of WP3 is to develop prediction and stratification tools based on multimodal data and ML/AI approach with clinical useful application. Specific objectives are to:

  • Build a pipeline for the integration of available data to facilitate algorithm development.
  • Ensure that the pipeline is flexible to allow for sample addition over the course of the project.
  • Develop algorithms for stratification of treatment outcome (response, adverse effects).
  • Develop prediction algorithms for selection of optimal treatment with minimal adverse effects.
  • Perform in-sample validation (sensitivity and specificity analysis) of the proposed algorithms.

Description of work

WP3 will apply the infrastructure developed by WP1 and develop specific artificial intelligence (AI) and machine learning (ML) tools to predict symptom severity and the development of adverse effects and multimorbidities in mental illness, and tools to stratify patients based on different treatment trajectories. WP3 will obtain randomised controlled trial (RCT) drug data, prescription registry data and data from eHRs from WP1, as well as related adverse effects of antipsychotics/antidepressants/mood stabilisers. These data will be analysed and integrated with genetic data from WP2 to develop prediction/stratification algorithms suitable for individualised

treatment (precision medicine), and ensure that the tools are developed for clinical applications.

The WP will define the mental disorder status of already genotyped cases (cross-referenced, nation-wide biobanks), as well as their treatment response status and development of adverse effects or multimorbidities, based on information from: i) eHRs; ii) registries or iii) health surveys/questionnaires (inference). Information from eHRS/registries will enable life-course observation through recorded ICD diagnoses in in- and outpatient clinics or inference through medication use from prescription registries. The latter two will allow us to maximise the numbers of cases. We estimate that we will have mental phenotypes from n=260k.

The WP will perform country-specific collection and combine and analyse data in the secure infrastructure built up from Tryggve (WP1). We will revise and extend the PHS method to two proof-of-principle applications: i) stratification of treatment outcome (response, adverse effects), building on the unique longitudinal outcome data (drug response, disease outcome) from registries and eHealth records, ii) prediction of onset of adverse effects or multimorbidities, including shared polygenic factors across traits and non-genetic data (16) (environmental factors, life events and self-report information). The inclusion of baseline data (previous episodes, sociodemographic situation) to improve the prediction accuracy is a critical step towards clinical utility. We will target highly relevant clinical scenarios, such as CVD after a depressive or psychotic episode above 60 years of age.

We will develop a tool to assist in the prediction of multimorbidities in severe mental disorders, which we can then extend to multimorbidity prediction after specific life events. Applying polygenic stratification and prediction tools to the treatment of mental disorders has not yet been possible because of the lack of large enough training sets and appropriate analytical tools to enable efficient exploration of the big data. Both the necessary samples and novel ML tools are available to the REALMENT multidisciplinary team.

Key WP3 tasks

  • Building a flexible pipeline for the integration of data to facilitate algorithm development.
  • Development of a novel tools for accurate stratification of treatment trajectories.
  • Development of a novel prediction tools for precision psychiatry applications.
Published Jan. 25, 2022 11:17 AM - Last modified Jan. 25, 2022 11:17 AM