1
Data quality over algorithm sophistication. 60% of project time went to cleaning and feature engineering — early models trained on uncleaned data failed despite algorithm sophistication.
← Applications
3.2 · Applications
Case Studies
Theory illuminates possibility, but case studies prove viability. Three implementations — hospital readmissions, real-time fraud detection, federated predictive maintenance — each with challenge, approach, results, and key learnings.
£5.8M
Annual revenue protected — hospital readmissions
0%
Fraud losses reduced — digital bank
0%
Unplanned downtime cut — manufacturing consortium
Case Study 1 · Healthcare
Regional Hospital Network — Reducing Readmission Rates
The Challenge
A 15-facility network faced 30-day readmission rates four percentage points above national averages — costing ~£8M annually in lost revenue. Traditional risk scoring relied on clinical intuition, failing to target which patients needed intensive discharge planning.
- Two-analyst internal team with limited ML expertise
- Budget constraints precluded commercial solutions
- Must integrate with existing EHR without new infrastructure
The Approach
- 5 years of records: 40,000+ discharge episodes
- Features: demographics, labs, vitals, meds, social determinants (transport, support)
- 60% of project time on data cleaning, validation, feature engineering
- Trajectory features: did conditions improve or worsen during admission?
- Deployed logistic regression — interpretability prioritised over marginal accuracy gain
- Risk tiers: low / moderate / high — mapped to standard / enhanced / intensive discharge
- Scores embedded directly into EHR discharge module; alerts fire at discharge eligibility
- Unit champions, monthly reviews, feedback loops for continuous improvement
−3.2 pts
30-day readmission rate
0
readmissions avoided per year
£5.8M
revenue protected annually
0%
high-risk patients identified
| Metric | Outcome |
|---|---|
| 30-day readmission rate | Down 3.2 pp (below national average) |
| Readmissions avoided / year | ~400 |
| Revenue protected / year | ~£5.8M |
| High-risk patients identified | 72% |
| Data cleaning share of project | 60% |
Key Learnings
2
Interpretability matters in healthcare. Feature importance visualisations and individual prediction explanations built trust even when predictions contradicted clinical intuition.
3
Integration makes or breaks implementation. A standalone dashboard with separate login had negligible usage. Embedding scores into existing EHR workflows transformed adoption — placement, timing, and report format all mattered.
4
Change management is underrated. Frame predictions as decision support rather than decision replacement. Unit champions provided peer credibility that formal training alone could not achieve.
5
Monitor post-deployment. Performance degraded as patient populations shifted. Quarterly validation and retraining on recent data proved essential — the original model would have lost effectiveness without ongoing attention.
Case Study 2 · Financial Services
Digital Bank — Real-Time Fraud Detection at Scale
The Challenge
A digital-first bank processing 20M+ transactions monthly faced £15M annual fraud losses. Rule-based systems declined ~5% of legitimate transactions, frustrating customers and driving attrition.
- Detection latency must stay under 100 ms
- Explainability required for regulatory audits and customer disputes
- Geographic expansion into new markets defeated existing rules
- Synthetic identity fraud and account takeover exploited rule-based gaps
The Approach
- Ensemble: gradient boosting + neural networks (sequence modelling) + graph algorithms (account networks)
- Velocity, geographic, device-fingerprinting, and behavioural pattern features
- Sequence models detect small test purchases before large fraud attempts
- Class imbalance (<0.1% fraud): custom loss functions penalising missed fraud
- 3-month shadow-mode deployment before cutover — validated without affecting customers
- Feature store pre-computed expensive features; GPU acceleration for neural components
- Hybrid explainability: interpretable secondary models approximate ensemble logic for declines
0%
fraud losses reduced
1.8%
false positive rate (was 5%)
73 ms
average processing latency
0%
account takeover detection (was 42%)
| Metric | Before | After |
|---|---|---|
| Fraud losses | ~£15M / year | Down 68% (>£10M saved) |
| False positive rate | 5% | 1.8% |
| Processing latency | — | 73 ms average |
| Account takeover detection | 42% | 89% |
| Synthetic identity detection | Unaddressed | 76% |
| Manual review queue | Baseline | Down 70% |
Key Learnings
1
Real-time ML engineering is not model development. Latency constraints required optimising the entire pipeline — feature computation, inference, decision logic — not just the model itself.
2
Imbalanced datasets need specialised handling. Standard accuracy optimisation produces models that classify everything as non-fraud (99.9% "accuracy", zero usefulness). Optimise precision-recall curves with business-cost loss functions.
3
Shadow-mode deployment is invaluable. Three months running parallel to existing rules caught pipeline bugs and edge cases before cutover — far cheaper than a failed production launch.
4
Explainability and performance require hybrid thinking. Interpretable logistic regression lagged ensemble detection. Approximating ensemble logic with secondary interpretable models satisfied both regulatory and performance requirements.
5
Fraud evolves continuously. Automated retraining pipelines and model drift analysis are not optional — fraudsters adapt constantly, and models effective last quarter may already be stale.
Case Study 3 · Manufacturing
Manufacturing Consortium — Federated Predictive Maintenance
The Challenge
Eight automotive-component facilities faced ~£3.5M annual costs from unplanned downtime. Time-based preventive maintenance was wasteful; tribal knowledge was walking out with retiring technicians.
- Competitive concerns prevented direct data sharing between facilities
- Each site had separate maintenance systems with inconsistent data quality
- Limited maintenance windows meant false alarms disrupting production would kill adoption
The Approach
- Federated learning: local models trained per facility; only parameters shared centrally — data sovereignty preserved
- IoT sensors: vibration, temperature, pressure, power consumption, acoustic emissions
- Spectral analysis, wavelet transforms, and statistical summaries extract physically meaningful features
- RNNs model sequential sensor dependencies; anomaly detection flags departures from normal
- Survival analysis predicts remaining useful life for risk-based maintenance scheduling
- Validation: known-fault injection + retrospective analysis + cross-facility generalisation tests
- UX for technicians: green/yellow/red health indicators; automatic work-order generation; mobile diagnostic guidance
−42%
unplanned downtime
−33%
maintenance costs
0%
failures predicted in advance
9.3 days
average advance warning
| Metric | Outcome |
|---|---|
| Unplanned downtime | Down 42% |
| Production losses avoided / year | ~£1.5M |
| Maintenance costs | Down 33% |
| Failures predicted with advance notice | 78% |
| Average advance warning | 9.3 days |
Key Learnings
1
Industrial environments impose real-world constraints. Sensor failures, communication dropouts, and calibration drift require robust imputation and quality monitoring — clean data cannot be assumed.
2
Domain expertise is not external to technical work. Maintenance engineers identified physically meaningful sensor patterns that pure data-driven feature selection missed, and their judgement validated prediction plausibility.
3
Manufacturing cultures prioritise proven methods. Adoption required pilot demonstrations, framing systems as decision support rather than automation, and emphasising downtime reduction over model accuracy.
4
Rare-event prediction needs different metrics. Standard accuracy is misleading at low failure rates. Focus on precision at operationally relevant recall levels and use cost-sensitive loss functions weighted to business impact.
5
Organisational culture determines technical success. Executive sponsorship, cross-functional collaboration (data science, maintenance, operations, IT), and early wins that build momentum proved as critical as any algorithm choice.
Key Takeaways
- Healthcare: 3.2 pp readmission reduction, £5.8M annual savings, 72% high-risk identification
- Finance: 68% fraud reduction, 1.8% false positive rate, 73 ms latency, 89% account-takeover detection
- Manufacturing: 42% downtime reduction, 33% maintenance cost savings, 9.3-day advance warning
- Data quality, stakeholder engagement, and change management matter more than algorithm choice
- Implementation excellence determines whether technical success becomes business value