Close
Harmony, our AI-data tool, was developed during Wellcome’s Mental Health Data Prize; it enables faster retrospective harmonisation of different questionnaires used in mental health research.
This project runs from June 2022 to June 2024 and is funded by a Wellcome Trust Mental Health Data Prize.
Background
Reviews have estimated that over 280 questionnaires have been used to measure depression and we see a similar pattern for assessing other mental health concepts. Such heterogeneity makes it difficult to compare studies and determine whether the results (e.g. efficacy of treatment) are an artefact of the questionnaire being used. One approach to addressing this is the harmonization of questionnaires; i.e. identifying similar questions that tap into the same symptom from different scales, and testing their measurement properties and equivalence empirically – thus enabling researchers to combine data from multiple datasets and also to compare findings across studies, even when different measures have been administered. The process of retrospective measurement harmonization involved many manual steps, where researchers screened pages of study-meta data and variable information to identify what two studies may have in common.
As part of Wellcome’s Mental Health Data Prize, our team have developed Harmony, a free-to-use online tool that allows researchers to harmonize data from different studies automatically and within seconds. Harmony users can simply upload the study meta-data (word, CSV or pdf format) they want to compare to our platform. Harmony then extracts the text information provided for the different studies and uses natural language processing to identify variables that are comparable across the datasets based on their semantic content and assigns an empirical ‘similarity score’ to pairs of variables.
Methodology
Harmony uses state-of-the art Artificial Intelligence (AI), specifically natural language processing (NLP), to identify variables that are comparable across the datasets based on their semantic content. We tested the underlying NLP algorithms and improved their performance and conducted various user testing sessions to design the platform.