Periodic Reporting for period 2 - FAITH (a Federated Artificial Intelligence solution for moniToring mental Health status after cancer treatment)
Berichtszeitraum: 2021-05-01 bis 2022-07-31
People with cancer are four times more likely to take their own lives than people without the disease, with depression being one of the most common comorbidities in cancer patients. For those that are not at risk of suicide they are still at an increased risk of mortality compared with non-depressed patients. Too often the symptoms of depression go unnoticed or are dismissed as being general symptoms of their cancer. Until their depression is eventually diagnosed and treated, these people have undergone tremendous suffering and experienced greatly reduced quality of life. ICT has provided new innovations that can assist patients with chronic illness, such as depression. This ranges from the use of devices that can communicate various types of physiological signals that can be evaluated remotely by a health care expert, to mobile apps looking to address a specific health problem or develop a good habit. Companies, however, are offering their services in exchange for personal information, participating in what Harvard Business School professor Shoshana Zuboff calls surveillance capitalism. This means companies can profit from stockpiling as much data as they can. Their devices are broadcasting data back to their headquarters for them to stockpile, bypassing the people who have generated it. In order to collect this data, the companies have to maintain security loopholes. “As long as companies are free to gather as much data about us as they possibly can,” security expert Bruce Schneier says, “they will not sufficiently secure our systems. As long as they can buy, sell, trade, and store that data, it’s at risk of being stolen. As long as they use it, we risk it being used against us.” Revealing our health information through our devices has continuing consequences for us in this industry built on exploiting our data.
In the past few years, machine learning has led to major breakthroughs in various areas, such as natural language processing, computer vision and speech recognition. Much of this success has been based on collecting huge amounts of data. While machine learning can benefit from this “big data” to generate state-of-the-art models, most healthcare data is hard to obtain due to legal, privacy, technical, and data-ownership challenges, especially among international institutions where HIPAA and GDPR concerns need to be addressed. An emerging solution to this problem is Federated Learning, a new approach to machine learning where the training data does not leave the users’ computer at all. Instead of sharing their data, the devices of the users computes weight updates themselves using their locally available data. It is a way of training a model without directly inspecting users’ data on a server. The goal of the FAITH project is to provide an ‘AI Angel’ that remotely analyses depression markers, using federated learning, to predict negative trends in their disease trajectory, giving their healthcare providers advanced warnings to allow for timely intervention. These markers are treated under several distinct categories; Activity, Sleep, Appetite, and Voice.
In the past few years, machine learning has led to major breakthroughs in various areas, such as natural language processing, computer vision and speech recognition. Much of this success has been based on collecting huge amounts of data. While machine learning can benefit from this “big data” to generate state-of-the-art models, most healthcare data is hard to obtain due to legal, privacy, technical, and data-ownership challenges, especially among international institutions where HIPAA and GDPR concerns need to be addressed. An emerging solution to this problem is Federated Learning, a new approach to machine learning where the training data does not leave the users’ computer at all. Instead of sharing their data, the devices of the users computes weight updates themselves using their locally available data. It is a way of training a model without directly inspecting users’ data on a server. The goal of the FAITH project is to provide an ‘AI Angel’ that remotely analyses depression markers, using federated learning, to predict negative trends in their disease trajectory, giving their healthcare providers advanced warnings to allow for timely intervention. These markers are treated under several distinct categories; Activity, Sleep, Appetite, and Voice.
The first reporting period has dealt with identifying the main stakeholders of the FAITH framework and their needs, and based on that defining the appropriate uses cases. Best practice approach towards architecture definition was followed, paying close attention to the foundations of this work e.g. defining how properly to capture functional and non- functional requirements, and how to visualise architectural decisions so they can evolve efficiently as needs be in the project. These decisions are all captured in the first iterations of the corresponding WP2 deliverables and will be refined over the coming iterations.
The need for two related, but distinct framework implementations emerged as the project matured; a framework that would best serve the needs of the FAITH trial in terms of data capture and analysis, and one that would deliver the results of that analysis using federated learning (FL). An analysis of FL techniques was led in WP4, along with review of many of the necessary, supporting tools and libraries. Fundamental of course is the FAITH mobile application. In this regard a core code project including libraries necessary was provided, alongside with UI integration procedures, while an integration plan was developed and has been put in force to monitor the integration activities that have been performed. The project partners have had in-depth weekly meetings to define the clinical trial protocol, the Case Report Form (CRF) and integrated statistical plan, which needs to be approved before starting the trial.
Additionally, this reporting period has seen the rigorous specification of variables as part of study protocol, app development to capture these variables, and trial specification to frame it all. In parallel, the initial versions of the WP4 deliverables have laid out our approach towards a federated framework, explainable AI, and model compression; all important components in terms of ultimately deploying federated models.
In terms of implementing a voice/NLP interface to gauge a user’s mental outlook, this objective is multi-faceted, and consequently the project has worked towards it in several paths: capturing voice input via the mobile application and offloading for vocal feature analysis to cater for the trial, experimenting with both on-device and browser-based approaches to the vocal feature analysis so that ultimately this objective could be achieved in a federated privacy-preserving manner, investigation of different methods for capturing the voice recording (based on methods used in the literature), investigating methods for using NLP to automate questionnaire delivery e.g. speech recognition using TensorFlow.js.
A significant outcome of this reporting period was the extent of the dissemination efforts. The FAITH Consortium engaged in dedicated healthcare stakeholder clustering activities, in addition to the standard dissemination and communication activities. These activities targeted the healthcare community, to collect feedback able to influence the direction FAITH should go in. FAITH is a member of the Health & Care cluster funded by the European Commission. In parallel FAITH formed the establishment of a “Cancer Survivorship – AI for Well-being” cluster, five EU-funded projects brought together by common research interests (mental health, well-being, depression, and patient support) and a common approach (a participatory research vision).
The need for two related, but distinct framework implementations emerged as the project matured; a framework that would best serve the needs of the FAITH trial in terms of data capture and analysis, and one that would deliver the results of that analysis using federated learning (FL). An analysis of FL techniques was led in WP4, along with review of many of the necessary, supporting tools and libraries. Fundamental of course is the FAITH mobile application. In this regard a core code project including libraries necessary was provided, alongside with UI integration procedures, while an integration plan was developed and has been put in force to monitor the integration activities that have been performed. The project partners have had in-depth weekly meetings to define the clinical trial protocol, the Case Report Form (CRF) and integrated statistical plan, which needs to be approved before starting the trial.
Additionally, this reporting period has seen the rigorous specification of variables as part of study protocol, app development to capture these variables, and trial specification to frame it all. In parallel, the initial versions of the WP4 deliverables have laid out our approach towards a federated framework, explainable AI, and model compression; all important components in terms of ultimately deploying federated models.
In terms of implementing a voice/NLP interface to gauge a user’s mental outlook, this objective is multi-faceted, and consequently the project has worked towards it in several paths: capturing voice input via the mobile application and offloading for vocal feature analysis to cater for the trial, experimenting with both on-device and browser-based approaches to the vocal feature analysis so that ultimately this objective could be achieved in a federated privacy-preserving manner, investigation of different methods for capturing the voice recording (based on methods used in the literature), investigating methods for using NLP to automate questionnaire delivery e.g. speech recognition using TensorFlow.js.
A significant outcome of this reporting period was the extent of the dissemination efforts. The FAITH Consortium engaged in dedicated healthcare stakeholder clustering activities, in addition to the standard dissemination and communication activities. These activities targeted the healthcare community, to collect feedback able to influence the direction FAITH should go in. FAITH is a member of the Health & Care cluster funded by the European Commission. In parallel FAITH formed the establishment of a “Cancer Survivorship – AI for Well-being” cluster, five EU-funded projects brought together by common research interests (mental health, well-being, depression, and patient support) and a common approach (a participatory research vision).
The main innovation in this reporting period, and one with genuine potential impact, has been the development of the FAITH mobile application. It is under active development and will support cancer survivors by using AI services to analyse their activity, sleep, eating characteristics, and vocal changes. The primary outcome of the trial is to assess the effectiveness of federated learning in identifying and predicting depressive symptoms in cancer survivor patients. Our hypothesis is that the federated learning model will significantly detect depressive symptoms in cancer survivors. This is our central expected result, and the societal implications if this can be realised are profound.