Periodic Reporting for period 1 - AI4HOPE (ARTIFICIAL INTELLIGENCE BASED HEALTH, OPTIMISM, PURPOSE, AND ENDURANCE IN PALLIATIVE CARE FOR DEMENTIA)
Período documentado: 2024-01-01 hasta 2025-06-30
Dementia is a major global health concern affecting over 55 million people and expected to almost triple by 2050. AI4HOPE addresses unmet needs of people with dementia (PwD) and carers by integrating early palliative care (PC) from diagnosis with evidence-based, non-pharmacological digital interventions. It combines advance care planning (ACP) with multisensory emotion regulation and virtual reality supported pain management to improve quality of life (QoL), comfort and carer support. Drawing on proven early-PC models in oncology, AI4HOPE reframes dementia as a life-limiting condition requiring proactive, holistic supportive care across the disease trajectory, not only at end-of-life. This closes the gap between evidence and practice and shifts care from reactive symptom control to preventive, need-led management. The project aligns with EU and Alzheimer Europe priorities for equitable, accessible, affordable and inclusive care. Social sciences and humanities are embedded to ensure ethical, culturally sensitive and inclusive design, to co-create with PwD, families and professionals, and to shape governance for trustworthy, transparent AI. Pathway to impact: (1) person- and carer-level gains, reduced distress, better symptom relief and stronger decision-making; (2) system-level gains, earlier planning, reduced avoidable healthcare use and cost-effective personalisation of care. Digital health-literacy tools will improve public understanding, reduce stigma and support informed choice. Combining patient-centred methods, ethical AI and multidisciplinary collaboration, AI4HOPE will deliver sustainable, resilient care models for millions across Europe and beyond.
Explainable AI (XAI) for Pain and Anxiety Detection: Developed ensemble and deep-learning models, trained on dementia-related datasets, to detect pain, anxiety and emotion. Focused on XAI to ensure clinical interpretability. Utilized multimodal datasets (Wearable Stress and Affect Detection (WESAD), Psychophysiology of Positive and Negative Emotions (POPANE), Fudan Sleep, Lifesnaps Fitbit, Pain E-motion Faces) for emotion, stress, pain, and sleep analysis. Selected the lightweight Polar 360 wearable for real-world dementia care data collection.
Multilingual service to extract medical concepts from unstructured text: Initial pipeline being integrated into Patient Sensing Network. The service extracts PROBLEM, TEST, or TREATMENT concepts from an input string, not limited to clinical semantics, and is being mapped to Fast Healthcare Interoperability Resources (FHIR) Composition and augmented with Logical Observation Identifiers Names and Codes (LOINC) methodology.
Pipelines to extract observable cues of pain and distress from diary recordings: The end-to-end pipeline serves as a comprehensive system for automated pain and distress detection from patient diary recordings. The initial pipeline for symptoms of mental distress was adapted from the SMILE project and integrated into the AI4HOPE framework. The work carried out in the context of AI4HOPE includes adaptation of the observable to the context of PwD and implementation of models to extract pain digital biomarkers-(Activity continues).
Conversational agents' framework for patient engagement: Implemented a conversational framework using RASA (open-source conversational AI platform) based chatbots for collecting patient-reported outcomes and assessing distress/pain risks, with large language model-based assistants for personalized education and information delivery on dementia progression and QoL management. The operational RASA chatbot and initial prototype AI assistant will undergo patient-expert validation in the next sprint.
Multisensory Patient Sensing Network: Designed a multisensory emotion recognition and pain/distress assessment framework, comprising individual analytical pipelines and integrated services.
The Dementia Journey Companion (DJC): A digital assistant empowering PwD by documenting personal preferences and routines, preserving life stories, and providing tailored education through AI.
User Stories & Use Cases: These outline the initial processes and findings from stakeholder engagement and requirements gathering efforts within the AI4HOPE project. It details the methodologies adopted, including co-creation workshops, focus groups, and consultations with public involvement groups, emphasizing the integration of value-sensitive design within an agile development framework. It shows how the insights derived from the WP timeline are integrated into the project's evaluation strategy. We also define five personas, patients (mild–moderate dementia), informal carers (family/friends), formal carers (professional carers/health aides), nurses, and clinicians/healthcare professionals, and two pilot-aligned use cases: (1) Assessment & Monitoring (Pilot 1); (2) Living with Dementia, Journey & ACP (Pilot 2). Over 120 user stories span system components, capturing diverse desires, interactions, and requirements to guide the development and evaluation of AI-driven solutions for dementia care.
Feasibility testing: Two pilots (n=150 each) in mild PwD across Ireland, Germany, Portugal, Slovenia, Spain and the UK will assess feasibility and acceptability of the digital intervention and the DJC, measuring perceived benefits/burdens and QoL impact.
Multilingual service to extract medical concepts from unstructured text: Initial pipeline being integrated into Patient Sensing Network. The service extracts PROBLEM, TEST, or TREATMENT concepts from an input string, not limited to clinical semantics, and is being mapped to Fast Healthcare Interoperability Resources (FHIR) Composition and augmented with Logical Observation Identifiers Names and Codes (LOINC) methodology.
Pipelines to extract observable cues of pain and distress from diary recordings: The end-to-end pipeline serves as a comprehensive system for automated pain and distress detection from patient diary recordings. The initial pipeline for symptoms of mental distress was adapted from the SMILE project and integrated into the AI4HOPE framework. The work carried out in the context of AI4HOPE includes adaptation of the observable to the context of PwD and implementation of models to extract pain digital biomarkers-(Activity continues).
Conversational agents' framework for patient engagement: Implemented a conversational framework using RASA (open-source conversational AI platform) based chatbots for collecting patient-reported outcomes and assessing distress/pain risks, with large language model-based assistants for personalized education and information delivery on dementia progression and QoL management. The operational RASA chatbot and initial prototype AI assistant will undergo patient-expert validation in the next sprint.
Multisensory Patient Sensing Network: Designed a multisensory emotion recognition and pain/distress assessment framework, comprising individual analytical pipelines and integrated services.
The Dementia Journey Companion (DJC): A digital assistant empowering PwD by documenting personal preferences and routines, preserving life stories, and providing tailored education through AI.
User Stories & Use Cases: These outline the initial processes and findings from stakeholder engagement and requirements gathering efforts within the AI4HOPE project. It details the methodologies adopted, including co-creation workshops, focus groups, and consultations with public involvement groups, emphasizing the integration of value-sensitive design within an agile development framework. It shows how the insights derived from the WP timeline are integrated into the project's evaluation strategy. We also define five personas, patients (mild–moderate dementia), informal carers (family/friends), formal carers (professional carers/health aides), nurses, and clinicians/healthcare professionals, and two pilot-aligned use cases: (1) Assessment & Monitoring (Pilot 1); (2) Living with Dementia, Journey & ACP (Pilot 2). Over 120 user stories span system components, capturing diverse desires, interactions, and requirements to guide the development and evaluation of AI-driven solutions for dementia care.
Feasibility testing: Two pilots (n=150 each) in mild PwD across Ireland, Germany, Portugal, Slovenia, Spain and the UK will assess feasibility and acceptability of the digital intervention and the DJC, measuring perceived benefits/burdens and QoL impact.
Preliminary Results: Our personalized model reached 95.59% accuracy on POPANE, outperforming prior results. On WESAD, a multistage framework with hierarchical feature fusion and voting ensembles yielded ~9% improvement over the original study. Next steps: large-scale validation on AI4HOPE data, real-world testing, commercialization/intellectual property rights, and ethical/regulatory compliance. DJC goes beyond static ACP tools by embedding early, comprehensive PC across the full dementia trajectory, not only at end of life. Drawing on the proven oncology model, it offers dynamic, stage-appropriate support for symptoms, psychosocial and spiritual needs, and continuous QoL optimization. DJC provides dynamic, stage-appropriate support and is co-created with stakeholders, including PwD, ensuring real-world needs drive its design. The weakly supervised multilingual medical Named Entity Recognition framework addresses critical limitations in current state-of-the-art models by extending medical named entity recognition to low resource where no existing models are available, while also bridging the gap between formal clinical language and informal patient expressions. Unlike existing models limited to single languages or reliant on extensive manual annotations, this framework employs a novel pipeline combining automated annotation with the high-performing Stanza i2b2 (pre-trained clinical extraction model) (F1 score: 88.1%), machine translation, word alignment, and Bidirectional Encoder Representations from Transformers fine-tuning to achieve robust multilingual performance without costly manual labelling. It significantly outperforms existing baselines, improving F1 scores from 67.7%→80.1% in English, 57.1%→75.6% in Italian, and 62.6%→77.6% in Spanish. Additionally, it effectively extracts symptoms from natural, conversational patient language, extending beyond traditional formal clinical terminology.