Safe Drive Africa: PhD Research and Continuing Impact

Safe Drive Africa is a completed PhD research programme and continuing impact initiative focused on reducing unsafe driving in Nigeria through culturally grounded, AI-enabled mobile feedback. The work combines smartphone telematics, machine learning, natural language generation, and behaviour-change theory to create practical, low-cost interventions for road safety in low-resource environments.

The programme originated as doctoral research at the University of Aberdeen addressing a major gap in road safety technology: most telematics-based behaviour-change systems were designed for Western, high-income settings and did not transfer well to countries like Nigeria. The programme responded by designing an intervention grounded in Nigerian realities — including weak insurance-based telematics infrastructure, inconsistent internet connectivity, limited speed-limit data, local driving culture, and the importance of legal authority in persuasive communication.

The result was an AI-enabled, culturally attuned mobile system that detects unsafe driving behaviours, generates persuasive feedback grounded in Nigerian traffic law, and supports safer driving through legally grounded, supportive communication rather than financial incentives. A within-subjects evaluation with 54 drivers demonstrated statistically significant behavioural improvement: Unsafe Behaviours Per Kilometre (UBPK) decreased 19.4% (1.44 to 1.16 events/km, p < 0.001), with 79.6% of drivers improving.

What began as a doctoral project now forms the foundation for continuing work on institutional adoption, fleet deployment, and scalable road safety solutions for West Africa.

Telematics feedback systems are often developed in high-income settings with different road infrastructure, regulatory environments, data availability and user expectations. Safe Drive Africa investigated what changes when such an intervention is designed for a Nigerian deployment rather than assuming that an existing model can simply be transferred.

Local requirements and safety priorities. The Nigerian stakeholder study identified gaps in formal training and traffic-rule knowledge alongside locally salient risks, including alcohol-related driving concerns. This created a requirement for feedback that could educate as well as coach.

Infrastructure and connectivity. Deployment experience showed the importance of local persistence because mobile connectivity could be intermittent and data costs could make always-online operation impractical.

Evidence availability. Smartphone sensors could support the measurement of harsh braking, rapid acceleration and swerving, but reliable machine-readable posted speed-limit information was not consistently available. The final Nigerian evaluation therefore excluded speeding from its UBPK outcome rather than treating GPS speed alone as proof of a legal speeding violation.

The resulting design combined authoritative legal grounding with supportive behavioural coaching while limiting claims to behaviours that could be measured credibly.

The final field evaluation addressed three questions:

RQ1. Does feedback from the AI-enabled mobile application reduce Unsafe Behaviours Per Kilometre when intervention driving is compared with the pre-intervention period?

RQ2. How consistently do drivers maintain improvement during the intervention?

RQ3. Do drivers who use the application more frequently during the intervention show larger reductions in unsafe behaviour?

The broader doctoral programme also investigated stakeholder perceptions of unsafe driving in Nigeria, the development of smartphone-based alcohol-influenced driving classification, and the design of culturally attuned AI/NLG feedback.

Research objectives: 1. Identify stakeholder concerns about unsafe driving and technology-supported road-safety interventions in Nigeria.

2. Collect real-world smartphone-sensor driving data.

3. Develop and evaluate machine-learning approaches for alcohol-influenced driving classification.

4. Design an offline-first mobile system for unsafe-driving analysis.

5. Develop legally grounded and persuasive Natural Language Generation for driver feedback.

6. Evaluate the intervention in a real-world Nigerian field study.

Stakeholder Requirements Study

A mixed-methods survey recruited 80 participants using stratified sampling: 51 drivers and 29 FRSC officials. Two role-specific questionnaires combined quantitative and open-ended questions.

Open-ended responses were analysed inductively using NVivo, while quantitative items were analysed using descriptive statistics and Fisher’s exact tests. Ethical approval was obtained from the University of Aberdeen Physical Sciences and Engineering Ethics Board, and participation was voluntary with informed consent.

Smartphone Data Collection

The research used Android smartphones to collect accelerometer, gyroscope and GPS data during real-world driving.

Development progressed through an initial telematics data-collection application and the later integrated Safe Drive Africa application. The final system used local data persistence and an unsafe-behaviour analyser with adaptive/context-sensitive detection rules.

The principal final evaluation metric, Unsafe Behaviours Per Kilometre, included harsh braking, rapid acceleration and swerving.

Speeding was excluded from the final UBPK outcome because reliable posted speed-limit reference data could not be consistently obtained.

Alcohol-Influenced Driving Classification

The research developed alcohol-classification models in two iterations.

The first modelling stage used Phase-1 driving data and produced an interpretable Decision Tree Pilot Model. That model was subsequently deployed on-device in the Safe Drive Africa application and evaluated against available driver self-reports during the later deployment.

A second modelling stage combined labelled data from the two collection phases and evaluated several models using a more rigorous nested cross-validation pipeline. The best-performing final offline benchmark was a Bagging Decision Tree model, which achieved:

AUC: 0.855, 95% CI 0.794–0.909;

mean F1: 0.589; and

mean recall: 0.615.

This Final Benchmark Model was not redeployed during the field intervention.

The classifier should be interpreted as detecting trip-level smartphone-sensor patterns associated with self-reported alcohol consumption. It is not a breathalyser, blood-alcohol measurement or legal determination of impairment.

AI/NLG Feedback

Safe Drive Africa used a dual-component feedback architecture.

Legally Grounded Tips. These were short, event-based messages generated from detected unsafe behaviours. Relevant Nigerian legal material was first retrieved from a curated knowledge base built from the Nigerian Highway Code and National Road Traffic Regulations. Prompt controls required laws, fines and penalties to come from that supplied context rather than being invented by the language model.

Tips were presented once daily from detected driving events rather than continuously while a driver was driving.

Weekly Reports. These transformed structured driving statistics into a supportive personalised narrative. The thesis implementation used a two-stage process: an initial GPT-4-turbo draft followed by a mandatory GPT-4o-mini refinement step designed to improve factual consistency, coherence, tone and implicit Theory of Planned Behaviour coverage.

The important research contribution is therefore the separation of legal grounding from persuasive behavioural reflection, rather than the use of any particular language-model provider.

Field Evaluation

The final application evaluation used a within-subject pre-intervention/intervention design.

The pre-intervention period ran from 10 January to 15 April 2025, during which feedback was disabled. The intervention period ran from 20 April to 15 July 2025, during which drivers received legally grounded Tips and weekly Reports.

Two transition weeks containing a mixture of pre-intervention and intervention trips were excluded from the comparison.

The final clean analysis comprised:

54 drivers;

402 pre-intervention trips;

400 intervention trips;

802 trips in total;

1,749.9 km pre-intervention; and

1,579.3 km during intervention.

A small follow-up qualitative component involved four semi-structured driver interviews.

Safe Drive Africa became a modular Android application integrating:

The architecture used local-first persistence so core data collection and analysis could continue when connectivity was unavailable.

Its main scientific feedback contribution was the separation of two complementary user-facing outputs:

Tips: concise, legally grounded, event-based educational feedback.

Reports: weekly supportive narratives designed to help drivers reflect on behavioural patterns, perceived control and practical actions for safer driving.

Overall Unsafe-Behaviour Reduction

Mean driver-level UBPK decreased from 1.44 events/km before feedback to 1.16 events/km during the intervention. This represents a 19.4% reduction.

The paired comparison was statistically significant:

paired t-test: p < .001;

Wilcoxon signed-rank test: p < .001; and

paired Cohen’s d magnitude: 0.660, a medium effect.

43 of 54 drivers (79.6%) improved.

These findings apply to the combined UBPK measure based on harsh braking, rapid acceleration and swerving. The final research does not report a validated 32.5% reduction in rapid acceleration, a 19.8% reduction in swerving, or a 10.2% reduction in harsh braking, and those figures should not be used as final study findings.

Consistency of Improvement

The research also showed that drivers reaching similar final UBPK levels did not necessarily improve in the same way.

Some showed steady improvement with few or no relapses, while others reached similar final outcomes through more volatile patterns with repeated higher-risk trips.

This finding motivated the use of bad-day frequency, based on how often an intervention trip exceeded a driver-specific pre-intervention risk threshold.

Engagement

The number of intervention trips had only a negligible overall association with improvement:

Pearson r = −0.071;

Spearman ρ = −0.043.

This means greater app use alone did not systematically predict greater improvement.

The 6–8 intervention-trip group contained the highest descriptive proportion of improvers, with 22 of 24 drivers (91.7%) improving, but this should not be interpreted as evidence that 6–8 trips is an optimal intervention dose.

What Participants Said About the Feedback

The four follow-up interviews provide qualitative context rather than population-wide causal evidence.

Legal Grounding Added Credibility

Participants reported that references to laws, fines and penalties made the feedback feel authoritative and consequential.

“The fines that are attached to it! It makes you sit up. You know that this is not a joke.”

Another participant emphasised that seeing a specific regulation made the feedback feel like more than generic app advice.

Supportive Tone Increased Acceptability

Participants also valued the encouraging tone of the weekly Reports.

“The tone of the report is very encouraging. It’s not judging... it’s not trying to condemn you, rather it’s telling you how you can improve.”

Tips and Reports Served Different Purposes

Participants did not consistently prefer one feedback type over the other. Instead, they described the two as complementary.

Tips supported short, actionable correction and education, while Reports provided a broader weekly opportunity for reflection.

Because only four drivers participated in these interviews, these findings should be treated as qualitative explanations of participant experience rather than generalisations to all Nigerian drivers.

The programme's effectiveness came from combining behavioural theory with local adaptation. The system used legal references from Nigerian road traffic regulations to strengthen credibility and deterrence, while maintaining a supportive, non-judgemental tone that kept drivers engaged rather than alienated.

Legal grounding enhanced credibility: Explicit references to Nigerian laws, fines, and penalties made feedback authoritative and consequential. As one driver emphasized: 'The fines that are attached to it! It makes you sit up. You know that this is not a joke.'

Supportive tone increased acceptability: The non-judgemental, encouraging tone made feedback motivating rather than alienating. Drivers praised: 'The tone of the report is very encouraging. It's not judging... it's not trying to condemn you, rather it's telling you how you can improve.'

Complementary roles of Tips and Reports: Drivers valued both feedback types for different purposes. Tips provided immediate, actionable corrections; Reports enabled weekly reflection.

The project demonstrated that effective digital road-safety support in Nigeria does not need to depend on insurance incentives or always-on connectivity. Instead, context-aware, offline-first, culturally grounded design can produce meaningful change.

The programme produced several software tools and datasets:

The programme has produced peer-reviewed publications, preprints, conference papers, presentations, working software, and datasets contributing to road safety, smartphone telematics, machine learning, and natural language generation.

Welcome to the Safe Drive Africa research team! Our dedicated team is composed of accomplished researchers, academics, and professionals who are passionate about leveraging technology to promote safe driving behaviours and enhance road safety in Nigeria and beyond. Get to know our team members and their invaluable contributions to this innovative research project.

Ethics and Privacy: Ethics and participant privacy were central to the programme from the beginning. Ethical approval was granted by the University of Aberdeen Physical Sciences and Engineering Ethics Board. All participants provided informed consent, data were anonymised for analysis, and drivers received £30 compensation. The system was designed with privacy-aware handling of behavioural data throughout.

Funding: The programme was supported by the Tertiary Education Trust Fund (TETFund) — a distinguished organisation established by the Nigerian government to drive the advancement of higher education, research, and development. TETFund's generous funding empowered the programme to pioneer innovative approaches for tackling road safety in Nigeria.

Additional support came through the University of Aberdeen's research infrastructure and, post-PhD, through IEAF (Impact, Engagement and Agile Funding) for commercialisation exploration activities.

Safe Drive Africa is no longer just a doctoral project. It now stands as the research foundation for a continuing programme of work on AI-enabled, low-cost road-safety interventions for Nigeria and comparable contexts.

The completed PhD established the conceptual, technical, and evaluative foundation. It demonstrated that culturally adapted persuasive technology can achieve meaningful behaviour change in developing countries without reliance on financial incentives or insurance infrastructure. It showed that ML-based alcohol detection from smartphone sensors can operate effectively in real-world LMIC contexts where breathalyser testing is impractical. And it provided a template for adapting Western telematics interventions to non-Western LMIC contexts through explicit local legal grounding and cultural alignment.

The next stage is about extending reach, adoption, partnerships, and real-world implementation. Future directions include ethical RCTs, broader deployment across African regions, adaptive/personalised feedback, integration of speed-limit data as it matures, partnerships with fleets and insurers, and cost-effectiveness analyses for policymakers.

This makes the programme section an important bridge within the personal website: it connects completed doctoral scholarship to ongoing work in applied AI, telematics, persuasive feedback, institutional engagement, and safer mobility in West Africa.