Algorithm Development for Training and Optimisation
Pubrica designs, trains, validates and optimises custom algorithms for healthcare, life sciences and pharmaceutical organisations. Our data scientists, biostatisticians and subject experts work with research groups, healthcare providers, pharmaceutical, biotech and medical device companies, and food and nutraceutical organisations that hold large, complex datasets under privacy and regulatory constraints. We develop AI, machine learning, deep learning and optimisation algorithms that automate workflows and support discovery, and we deliver the validated model with its code, environment, documentation, a benchmark against a baseline and a monitoring plan.
- Domain-specific development: we work in healthcare, life sciences, pharma, medical devices, food and nutraceuticals, where data constraints and failure modes differ by field
- Validation against held-out data: we apply correct cross-validation and report out-of-sample performance
- Optimisation for your objective: we balance accuracy against latency, interpretability, cost and calibration, as most deployments require
- Privacy and compliance from the start: HIPAA, GDPR and FDA or EMA expectations determine what can be built before any model is trained
- Post-deployment support: we monitor models in production, because performance drifts as the population and recording practices change
Design, training and optimisationIncluded
Validation against held-out dataIncluded
Documentation, code and benchmarksIncluded
Deployment and integrationIncluded
Regulatory clearance of the modelClient responsibility
Accuracy promised in advanceCannot be promised
Problem → designed → trained → validated → deployed
- Baseline established before anything complex
- Validated on data it has not seen
- Objective and trade-offs agreed first
- Code, environment and documentation delivered
- Monitoring plan for after deployment
Service Scope
Six Components of Every Algorithm Development Engagement
Each engagement is scoped in writing, with the objective, the trade-offs and the success criteria agreed before development begins.
From the baseline reportDeep learning proposed for the classification taskA logistic regression on six features reaches 0.79 AUC. The neural network reaches 0.81, needs a GPU and cannot be explained to a regulator. The report recommends the regression and shows the comparison.
From the design noteObjective: flag for review, not decide. That changes the metric from accuracy to recall at a fixed review capacity, and it changes the model. Agreed before development, not after the first evaluation.
From the training logFeature selection on the full dataset, then nested CVSelection moved inside the outer folds. Reported performance dropped from 0.89 to 0.78, and 0.78 is the number that will hold.
From the validation reportPerformance by site: 0.81, 0.79, 0.80 and 0.62. The fourth site records the key variable differently. Reported rather than averaged away, because averaging it hides the one place the model fails.
From the handoverContainer with pinned dependencies · inference code and preprocessing as one artefact · model card with intended use, training population and known limitations · rollback procedure.
From the monitoring plan“Retrain annually.”Input distribution monitored monthly against the training distribution, performance tracked where labels arrive, and a defined trigger for retraining rather than a calendar date.
Samples are illustrative and anonymised. Hover any one of them to play it again.
Types of Algorithms
Eight Types of Algorithms We Develop
We select the approach based on the data type, the operational constraints and whether the output must be explainable.
Domains. Healthcare, life sciences, pharmaceuticals, medical devices, food and nutraceuticals. What we will not do: propose a complex model without comparing it to a simple baseline, report in-sample performance as performance, or build a clinical-decision algorithm without establishing first whether it is a regulated device.
Who we serve
Seven Client Groups We Support with Algorithm Development
Each client group has different requirements, whether the algorithm must satisfy a peer reviewer, a regulator, a production system or a clinician acting on its output.
Compliance and Standards
Algorithm Development Standards and Regulatory Frameworks
What governs an algorithm depends on what it is for. A research model and one that informs a clinical decision answer to very different requirements.
TRIPOD+AI (BMJ, 2024)
The reporting standard for clinical prediction models including machine learning ones, and the checklist a reviewer will apply.
Data protection (GDPR, HIPAA and local regimes)
Where patient data is involved, what may be processed and where is settled before design rather than documented after it.
Software as a medical device (FDA and EMA expectations)
Where the algorithm informs a clinical decision it may be a regulated device. That changes the evidence, the documentation and the change-control requirements substantially.
Information security management (ISO/IEC 27001:2022)
The current version of the information security standard. The 2013 edition has been superseded and its transition window has closed.
ICMJE Recommendations (January 2026)
Where the work is published, including declarations about data, code and AI tool use.
Good Clinical Practice (ICH E6(R3), January 2025)
For algorithms developed inside or alongside clinical trials.
Quality management (cert. H3CR9TCU20251212INDQM10N)
How the work is run, recorded and reviewed — not a claim about algorithm performance.
Responsible Use of AI
Explainable and Defensible Algorithm Development
In algorithm development, AI is the product rather than a supporting tool. Reviewers, regulators and clinical partners need to know why a model was chosen and how it behaves when it is wrong, so every modelling decision is documented and explainable.
Why We Do Not Publish an AI Accuracy Figure
Accuracy figures for AI tools are usually reported by the companies that sell them.
Pubrica therefore does not quote its own accuracy figure, because it would carry the same conflict of interest. Instead, we tell you which tool was used on your project, at which stage, and what a human reviewer changed afterwards.
Methodology
Our Seven-Step Algorithm Development Methodology
The process runs from feasibility assessment at step one to post-deployment monitoring at step seven.
Requirement analysis and problem definition
We define what the algorithm is for, what data exists, which constraints apply and what success looks like, and we establish current performance as the baseline.
Data collection and preprocessing
We source, clean and normalise structured and unstructured data, and we identify and report bias and imbalance rather than correcting them silently.
Algorithm design and development
We design candidate approaches and compare them against a baseline, giving full consideration to the simplest adequate method.
Training and optimisation
We train models on real data and tune them for the objective agreed at step one, with all fitting done inside the training folds.
Validation and testing
We apply cross-validation, held-out testing and stress testing across sites, time periods and subgroups, and report performance by stratum as well as in aggregate.
Deployment and integration
We integrate the model into your systems and pipelines with the environment pinned, the code documented and a rollback path defined.
Monitoring and support
We monitor input drift and performance where labels arrive, with a defined retraining trigger, because production data differs from training data.
Request a Free Algorithm Feasibility Assessment
We provide a written note on whether the problem is tractable with your data, a sensible baseline, the approach we would try first and the realistic performance ceiling. It is delivered within two business days at no charge, including when a simple rule would be sufficient.
Please tell us what happens today and how well it works. If an algorithm cannot outperform the existing rule, that is a useful finding, and it is far less costly to establish at the start.
Our experts
Your Algorithm Development Team
A named team of data scientists, statisticians and subject specialists stays with your project from scoping to delivery. The handover records who completed each stage, and your project coordinator introduces each member by name before work begins.
Reviews the scientific content for accuracy and confirms that each conclusion is supported by the evidence. Matched to your field rather than allocated from a rota.
Checks that statistical methods, tables and conclusions are consistent with the analysis plan and that the analysis can be reproduced.
Prepares, cleans and documents the data so that every step from raw data to final output can be traced and reproduced.
Independently checks code, tables and outputs against the analysis plan. Discrepancies are logged and resolved before delivery.
Sample work
Sample Work: Classification Model as Proposed and as Delivered
Step through the findings one at a time. The delivered model is simpler than the proposed one, performs within 0.02 AUC of it, and is suitable for clinical deployment.
As proposed. Deep learning classifier on 40,000 records across four sites.
- Neural network, 3 layers
- Feature selection on the full dataset
- Reported AUC 0.89
- Performance reported in aggregate
- No baseline established
- “Retrain annually”
As delivered. Simpler, validated by site, and deployable.
- Penalised logistic regression, 6 features — within 0.02 AUC of the network
- Selection moved inside the outer folds — 0.89 became 0.78
- AUC 0.78 out of sample, calibration slope 0.91
- By site: 0.81, 0.79, 0.80, 0.62 — site 4 codes the key variable differently
- Baseline: the existing rule scores 0.71. The model earns its place, narrowly
- Drift monitored monthly, retraining triggered by a defined threshold
Figures illustrative; the clinical task, record count and values are changed. Every one of these findings is one we see regularly.
Benefits
Benefits of Pubrica Algorithm Development
Each of these is checkable in your own code, report and monitoring. None is a promise of a performance figure, which is what the validation exists to establish.
Packages
Algorithm Development Service Packages
Each package is priced on the state of the data, whether the problem is single- or multi-modal, and whether deployment and monitoring are in scope. The feasibility assessment is free and comes first.
Foundational algorithm development
Startups, small research projects and proof-of-concept studies
Quoted per projectPriced on dataset size
Delivery 2–3 weeks
- ✓Requirement analysis and feasibility study
- ✓Baseline established before modelling
- ✓Algorithm design for structured data
- ✓Training and validation against held-out data
- ✓Performance report with recommendations
Suitable for establishing whether the problem justifies a larger engagement.
View a sample Request a quoteOptimised algorithm development
Research groups and organisations that need a validated model integrated with their data pipelines
Quoted per projectPriced on data complexity
Delivery 4–6 weeks
- ✓Comprehensive requirement gathering
- ✓Design for structured and unstructured data
- ✓Training, optimisation and nested cross-validation
- ✓Validation stratified by site, period and subgroup
- ✓Integration with your existing data pipelines
- ✓Benchmarking against the baseline, and documentation
Our most commonly selected package. Stratified validation often produces the most useful findings.
View a sample Request a quoteAdvanced and custom algorithm solutions
Large-scale research, enterprise applications and regulated settings
Quoted per projectPriced on scope and deployment
Delivery 6–10 weeks
- ✓Full consultation and strategic planning
- ✓Custom development for multi-modal data
- ✓Advanced training with AI and machine learning methods
- ✓Real-time testing and fine-tuning
- ✓Deployment-ready artefact with pinned environment and model card
- ✓Post-deployment monitoring and support
Where the algorithm informs a clinical decision, regulatory classification is established at scoping, because it determines the evidence and documentation required.
View a sample Request a quoteSend the problem and a data sample and we will assess feasibility free, including what a sensible baseline looks like. Ongoing monitoring is quoted against a schedule.
Quality Assurance
Our Service Commitments
Our service commitment, its limits and the remedy we offer are set out below.
The Pubrica guarantee
Performance Benchmarked Against a Defined Baseline
Every engagement establishes what the current approach achieves before anything is built, and reports the new model against it on data neither has seen.
01
Baseline established first
We measure current performance or the performance of a simple model, so that new results can be compared against it.
02
Stratified out-of-sample reporting
Results are reported by site, period and subgroup as well as in aggregate, because aggregate figures can hide a failure.
03
Simplest adequate model considered
Where an interpretable model performs within a meaningful margin, the comparison is in the report and the recommendation explains the choice.
04
Advice when a simple rule is sufficient
Some problems do not need a model, and we will tell you when that is the case.
Client Stories
Healthcare and life sciences researchers, data science teams, technology companies and organisations that need predictive analytics, automation or decision support inside a regulated or evidence-driven environment.
In each case, the output must be defensible to a reviewer, a regulator or a clinician, not only functional.
Supervised, unsupervised and reinforcement learning; predictive and prescriptive models; rules-based and heuristic algorithms; computer vision, natural language processing and speech recognition; and multi-objective optimisation.
The choice depends on the data, the constraints and whether the output must be explainable; the need for explainability often narrows the options most.
Against a baseline, on data it has not seen. We establish what the current approach achieves, then report the new model out of sample with preprocessing, feature selection and tuning all fitted inside the training folds.
Performance is reported by site, period and subgroup rather than only in aggregate, because an average conceals the place where it fails — and that is usually the finding that matters most.
Typically: requirement analysis and design, one to two weeks; development and initial testing, three to six weeks; optimisation and validation, two to four weeks; deployment and documentation, one to two weeks.
The variable that moves this most is data condition. A dataset that needs harmonising before anything can be trained adds more time than the modelling itself takes.
Yes, and we recommend it for any model that informs real decisions. Input distributions are monitored against the training distribution, performance is tracked where labels arrive, and retraining is triggered by a defined threshold.
A model degrades as the population and recording practices change, and its outputs can continue to look reasonable while this happens.
Software executes rules you specify. An algorithm of this kind learns from data, which means its behaviour depends on data you did not write and may not fully understand.
The practical consequence is that the work is mostly validation. In conventional software you test whether it does what you specified; here you are establishing what it does at all, and on which populations.
Yes, and the first question is whether what you are building is a regulated device. An algorithm that informs a clinical decision may be software as a medical device, which changes the evidence, the documentation and the change-control requirements substantially.
We establish that classification at scoping, because discovering it after development requires substantial rework.
Related services
Related Data Analytics Services
These related services prepare the data, frame the analytical question and report the results.
1DATA PREPARATIONthis comes first
AI & data preparationCleaning, annotation, documentation and leakage checks. BioinformaticsPipelines, reproducible workflows and interpretation.2PREDICTIVE MODELLINGoften enough on its own
Predictive analyticsForecasting and risk models, validated out of sample. BiostatisticsAnalysis plans, reproducible code and the reporting.3REPORTING AND PUBLICATIONthe reporting side
Interpretation & visualisationGuideline-compliant reporting and publication-ready figures. Scientific editingThe reporting-guideline check and the simulated review.Insights
Further Reading on Algorithm Development
Request a Free Algorithm Feasibility Assessment
Within two business days and at no charge, we will assess whether the problem is tractable with your data, a sensible baseline, the approach we would try first and the realistic performance ceiling, including when a simple rule would be sufficient.