Hubert Kenfack Ngankam

We propose an auto-encoder architecture (AuAE) for detecting subtle sequencing errors in human activity recognition, with potential applications to motor disorder monitoring in ambient environments. The architecture incorporates entropyregularized attention mechanisms and context suppression, inspired by computational studies of enhanced local feature processing observed in certain neurocognitive patterns. AuAE combines bidirectional GRU encoding with attention gating to model spatio-temporal 3D mesh sequences (SMPL-X format). Evaluated on two AMASS datasets (i.e., BMLrub and CMU datasets) with held-out synthetic anomalies (30 % contamination rate), AuAE achieved improvements over standard GRU-AE: anomaly $\mathbf{F}_{1}$-score increased from 0.01 to 0.20 on BMLrub and from 0.13 to 0.18 on CMU datasets. Ablation studies indicate that each architectural component contributes incrementally to anomaly detection performance.

Background: Major neurocognitive disorders (MNCDs) frequently lead to difficulties in performing activities. Several studies have shown that people living with an MNCD benefit from the use of learning optimization methods from cognitive rehabilitation, such as error-free learning, motor encoding, spaced retrieval, and fading, which promote the safe pursuit of their meaningful activities. However, while the principles of learning optimization methods are relatively straightforward, the personalized application of these methods to the specific situations encountered can be more difficult. Objective: The aim of this study was to describe the codevelopment process, including the validation of a web application called “Aide-Mémoire-Interactif (AMI),” a tool to help in personalizing learning optimization methods. Methods: To design the web application, participatory action research based on a codevelopment and validation process was carried out. The various stages leading to the codevelopment of the AMI web application were grouped into 3 phases: (1) assessment of caregivers’ needs (identification of situations frequently encountered by people living with an MNCD); (2) production of the first version of the AMI web application; and (3) validation of the AMI application. Individual interviews and workshops were conducted with 20 participants (caregivers, health and social service professionals, and community organization workers) to obtain a diversity of viewpoints and application contexts of use. Results: The AMI web application was developed through an iterative, participatory process involving caregivers and professionals. Participants identified daily situations requiring the use of learning optimization methods and provided feedback that guided successive refinements to the content, answer choices, navigation, and personalization features. The logical architecture was built according to the rationale underlying learning optimization methods. The final version enables users to generate tailored strategies for specific situations, access educational capsules, and save personalized strategies. Participants reported improved clarity, usability, and relevance. Conclusions: This participatory action research enabled the codevelopment and validation of the AMI web application to support the operationalization of learning optimization methods for people living with an MNCD by personalizing them to the situations encountered daily as well as to the underlying cognitive difficulties. By involving caregivers, professionals, and users, this process contributed to the development of a web application that meets user needs and their appreciation.

Le maintien à domicile des personnes âgées, essentiel pour leur santé cognitive et leur qualité de vie, exige une surveillance fine des activités de la vie quotidienne (AVQ), notamment l’hygiène personnelle. Les solutions de surveillance non intrusives actuelles sont limitées : elles peinent à reconnaître des activités complexes et de longue durée, manquent de granularité pour différencier une variation bénigne (fatigue) d’une anomalie pathologique (erreur séquentielle liée à un déclin cognitif), et échouent face aux occlusions prolongées (rideau de douche). Ce projet propose une solution basée sur l’utilisation de données 3D (nuages de points LiDAR), cruciales pour la préservation de l’intimité, tout en capturant la géométrie fine des mouvements.

Assistive Technologies for Cognition (ATCs), such as the Cognitive Orthosis for coOKing (COOK), offer support to individuals with traumatic brain injury by enhancing safety and independence. While the usability of COOK's client interface has been tested, the expert interface-used by occupational therapists to customize the interface to clients' needs-requires further study. We therefore aimed to: (1) Prioritize the modifications to be made to COOK's configuration interface; and (2) Describe the effectiveness and efficiency of the interface within a laboratory context. A Human-Centered Design (HCD) approach was used. Fourteen occupational therapists and master's students in occupational therapy participated. A convergent mixed-methods design was used. Data was collected through laboratory testing of the interface. Qualitative data was analyzed using deductive thematic analysis [1], and quantitative data with descriptive statistics. Our study highlighted the importance of balancing effectiveness and efficiency during technology design. Participants emphasized that the configuration process should ensure both a quality that allows customization to meet client needs and an ease of use to minimize the time required for the configuration of COOK. Overall, a good usability of COOK's configuration interface was demonstrated both qualitatively and quantitatively. Future studies will be needed to implement the technology in clinical practice and gather further insights on this interface for the design of further iterative improvements to meet the needs of occupational therapists and their clients.

BACKGROUND People living with major neurocognitive disorder (MNCD) may experience difficulties in carrying out meaningful activities. Learning-optimizing methods, building on the abilities preserved despite MNCD, can be beneficial in maintaining activity achievement for these individuals. Although these methods can be understood by caregivers, and they want to use them, adapting them to everyday situations is a challenge. Several web or mobile applications with various functions have shown benefits for people living with MNCD and their caregivers. However, none of these tools is designed to enable the appropriation of learning-optimizing methods and their use in everyday situations. OBJECTIVE The aim of this study was to co-develop a web application called Aide-Mémoire-Interactif (AMI), designed to support family caregivers in operationalizing methods that optimize learning according to the specific situations encountered on a daily basis and the cognitive difficulties of the person living with an MNCD. METHODS To design the AMI web application, the research team, in partnership with a community organization, caregivers and healthcare professionals, carried out participatory action research including: 1) assessment of caregivers' needs concerning situations frequently encountered by people living with an MNCD; 2) development of a logic algorithm based on these situations and cognitive difficulties, leading to the first version of the AMI application prototype; and 3) experimentation by caregivers for validation, leading to the final version of the AMI application. RESULTS The AMI application was designed to be easy to navigate. Based on Agile-Scrum methodology, it is user-centered to offer dynamic interaction with generated content, and customizable, since it is organized around modular and adaptable software components. Validation meetings with caregivers and healthcare professionals enabled us to improve several elements of the application (e.g., adjusting the wording of daily situations to facilitate comprehension). The final version of the AMI application features three main functions: 1) navigation to operationalize the methods optimizing learning for a specific situation, 2) viewing of video capsules describing the methods or 3) consultation of saved results from previous navigations. CONCLUSIONS The involvement of caregivers and healthcare professionals in the development of the AMI application has resulted in an application that stands out for its customization to each person's context, its user-friendliness and its flexible web architecture. The AMI application has the potential to be a practical and effective tool to support the daily use of learning-optimization methods.

The integrity of sensor datasets used in smart home applications is crucial for tasks like activity recognition and automation. We identify common validity issues such as event ordering errors, lifecycle inconsistencies, and data corruption, which are often overlooked but can significantly affect the reliability of analyses. We present a toolbox based on the BeepBeep stream processing library that enables efficient verification of 19 sanity checks on data streams. Our analysis of 15 publicly available smart home datasets collected by five different research teams reveals that most of them violate key assumptions about sensor behavior, emphasizing the need for pre-validation.