Projects: Usman Sani Dankoly

Project Overview

My project develops a universal taxonomy of 24-hour physical behaviour that is inclusive across cultures and adaptable to evolving technologies. Instead of beginning from device signals alone, I start with how people naturally describe their day, as discrete, time-bounded events with purpose and context, and integrate these narratives with sensor data. I build on αBET, an agnostic, event-based hierarchy that represents daily life as sequences of named events (with start time, duration, intensity, and context). To ensure the taxonomy does not reproduce WEIRD-science biases, I apply an intersectional lens, prioritising language and region, and curate global diary narratives (“A Day on Earth”) alongside publicly available single-day vlogs. Transformer models are then used to extract events from text and map them to αBET, creating a culturally aware, future-proof representation of 24-hour behaviour. Project 1 – αBET Framework Development Existing classifications of physical behaviour often rely on device-specific thresholds that miss the contextual richness of everyday life. To address this, I finalised αBET (Agnostic Behavioural Events Taxonomy) through a multi-round Delphi consensus (2014–2024) with international experts at ICAMPAM 2017, 2019 and 2024. αBET conceptualises daily activity as sequences of discrete events marked by transitions in context, activity and time, grounded in cognitive event-segmentation theory and Badiou’s concept of contextual “voids.” This project has two primary aims: 1. To establish a hierarchical, event-based model capable of describing any 24-hour day across populations and settings, independent of device or discipline. 2. To publish a standardised data model and nomenclature, including event name, start time, duration, intensity and contextual metadata, that enables full interoperability between narrative and sensor-based datasets.

First Study

αBET Framework Development: Existing classifications of physical behaviour often rely on device-specific thresholds that miss the contextual richness of everyday life. To address this, I finalised αBET (Agnostic Behavioural Events Taxonomy) through a multi-round Delphi consensus (2014–2024) with international experts at ICAMPAM 2017, 2019 and 2024. αBET conceptualises daily activity as sequences of discrete events marked by transitions in context, activity and time, grounded in cognitive event-segmentation theory and Badiou’s concept of contextual “voids.” This project has two primary aims: 1. To establish a hierarchical, event-based model capable of describing any 24-hour day across populations and settings, independent of device or discipline. 2. To publish a standardised data model and nomenclature, including event name, start time, duration, intensity and contextual metadata, that enables full interoperability between narrative and sensor-based datasets.

Second Study

Intersectional Sampling Scoping Review: Most physical-activity studies under-represent marginalised groups and rarely apply intersectionality at the sampling stage. This project systematically reviewed how health and behavioural research integrates intersectionality into sampling frameworks, not only recruitment. Using PRISMA-ScR methods across seven databases and AI-assisted screening, I identified 23 studies (from > 90 000 records) that illustrate both good practice and persistent barriers such as cost and rural under-representation. This project has two primary aims: 1. To synthesise and critically appraise intersectional sampling methods across quantitative and qualitative research, identifying strategies that successfully capture diverse social identities. 2. To develop a pragmatic intersectional sampling framework for global diary studies that uses language–region stratification, minimum stratum floors (20–30 diaries per subgroup), and adaptive oversampling to guarantee equitable representation while maintaining feasibility. The resulting framework underpins data collection in Project 3, ensuring that the taxonomy’s evidence base reflects linguistic and ecological diversity rather than convenience samples.

Third Study

Universal 24-Hour Taxonomy from Multilingual Diaries: To make αBET empirically grounded and culturally inclusive, I built a global narrative corpus called “A Day on Earth.” The initiative collects single-day diaries and day-in-the-life vlogs from over 190 countries, translated into 30 languages and curated under strict GDPR-compliant protocols. Early pilots confirmed the lack of large-scale public diary data, prompting original multilingual survey design and ethical web-scraping of daily-life videos. The corpus powers a transformer-based Named Entity Recognition (NER) pipeline that automatically detects event boundaries and temporal cues in free-text diaries. For instance, “I woke up at 7 am, made tea, and walked to work” becomes a structured sequence of αBET-labelled events (wake-up → prepare food → commute), each with start time, duration and context. The pilot BERT model achieved macro-F1 scores of 0.82–0.88, confirming that everyday narratives can yield structured behavioural data. This project has two primary aims: 1. To curate a multilingual, intersectionally stratified corpus of 24-hour diaries that capture routine diversity across languages, regions and cultures. 2. To design and validate a reproducible NLP pipeline that extracts, classifies and visualises event sequences, developed jointly with the Glasgow School of Art to create an interactive representation of daily life across the globe.

Fourth Study

Predictive Modelling of Event Sequences: Daily routines are not static; they evolve in patterned yet flexible ways. To capture this dynamic nature, I represent each diary as an ordered sequence of αBET-labelled events and apply sequence-based transformer architectures, inspired by life2vec, to model transitions and routine stability over time. Each event token encodes context, duration and intensity, enabling the model to learn both temporal order and behavioural meaning. During pre-training, the model learns the “language” of daily events through masked-token and sequence-ordering objectives; in fine-tuning, it predicts next-event probabilities or deviations from habitual patterns. Evaluation uses accuracy, F1 and temporal-ranking metrics to ensure robust and interpretable predictions. This project has two primary aims: 1. To encode and model 24-hour event sequences using transformer architectures that capture temporal and contextual dependencies across populations. 2. To build interpretable predictive models capable of forecasting routine changes, supporting personalised health analytics, rehabilitation planning and policy evaluation. By linking αBET semantics with advanced sequence learning, this project creates a predictive behavioural grammar that reveals how routines emerge, persist or shift in real-world contexts.