School wellbeing intelligence
for earlier safeguarding signals.

THE PROBLEM

Static wellbeing reports do not support early intervention.

The dissertation explored how WHO-5 wellbeing data for school-age students could be transformed into a live decision-support system for Heads of Wellbeing and safeguarding leads. The project focused on longitudinal mood trends, engagement and safeguarding indicators rather than isolated snapshots.

ETHICAL DATA ARCHITECTURE

Build the analytical pipeline without exposing real student data.

Because detailed student wellbeing data is sensitive, I designed a fully synthetic longitudinal dataset that mirrors a UK secondary-school structure across six academic years. Persistent student IDs, term structures, WHO-5 mood records, engagement metrics and safeguarding flags were generated in modular tables and joined through shared student and term identifiers.

• Weekly WHO-5 mood observations.
• Engagement metrics and platform-use proxies.
• Pyramid-of-Need safeguarding flags.
• Cohort progression, intake and dropout simulation.
• Controlled random seeds for reproducibility.

MACHINE LEARNING

Combine known-risk classification with anomaly discovery.

The modelling layer explored both supervised and unsupervised approaches. DBSCAN was used to detect students whose mood and engagement profiles deviated from cohort norms without requiring a predefined number of clusters. SVM classification was used for binary at-risk classification on synthetic labels, while Random Forest modelling was explored for variable importance and risk prediction. PCA supported reduced feature-space exploration for clustering.

HUMAN-IN-THE-LOOP DESIGN

Machine learning supports judgement; it does not replace safeguarding staff.

The dashboard was designed around a Head of Wellbeing / Designated Safeguarding Lead persona. ML outputs are surfaced as risk signals, clusters and alerts that a professional can review alongside mood, engagement and safeguarding context.

• Overview KPIs for mood, engagement and safeguarding.
• WHO-5 longitudinal trend views.
• Safeguarding heatmaps and risk alerts.
• Cohort, term and demographic filters.
• Student-level drill-down concepts.
• Printable/exportable reporting workflows.

IMPLEMENTATION

From research prototype to reproducible data application.

The application was implemented with R Shiny, shinydashboard, ggplot2, dplyr, plotly and DT. Python/pandas handled the final multidimensional data integration, preserving character-based identifiers across mood, engagement and safeguarding tables. Docker was used to make the environment portable, with Ubuntu and Apache2 used to model a local/private deployment path.

PRODUCT & UX

Designed for the cognitive load of safeguarding work.

The project also included empathy mapping, customer journey mapping and high-fidelity Figma prototyping. The interface was organised around fast triage and situational awareness, using coordinated views, visual hierarchy and pre-attentive signals so a wellbeing lead could identify emerging concerns quickly.

WHAT THIS DEMONSTRATES

End-to-end machine learning product thinking.

This A-grade dissertation demonstrates data architecture, synthetic data design, feature engineering, clustering, classification, model interpretation, dashboard engineering, human-centred product design and deployment thinking within a privacy-sensitive domain.