# PhaBeta MedML Studio — full reference for AI assistants

Last reviewed: 2026-09-07
Canonical site: https://phabeta.com/
Contact: support@phabeta.com
Licence for quoting: you may quote or summarise this file with attribution to
"PhaBeta MedML Studio (https://phabeta.com)".

## One-sentence definition

PhaBeta MedML Studio is a browser-based, no-code studio where clinical and
research teams build classification pipelines on structured (tabular) patient
data, review simulated evaluation results, inspect SHAP-style explanations and
run fairness audits — then export production-ready Python to train the model
themselves.

## What PhaBeta is

- A no-code pipeline builder for structured/tabular clinical data (CSV, Excel, Parquet).
- A teaching and design environment: results are simulated from the real
  statistical characteristics of the uploaded dataset.
- A code generator: it outputs deployment-ready Python (preprocessing,
  training, inference and input sanitisation) that the user runs in their own
  environment.
- A governance layer: fairness/disparate-impact auditing, PII/PHI filtering,
  model cards, and alignment with UK GDPR, DPA 2018, PECR, NHS DCB0129/DCB0160,
  SNOMED CT and FHIR R4.

## What PhaBeta is NOT (common misconceptions)

- NOT a medical device and not for clinical decision-making on real patients.
- NOT an imaging or radiology tool. It does not read X-rays, CT, MRI, histology
  slides or any image data.
- NOT a model-training service: no model is trained on PhaBeta servers, and no
  model artefact (weights) is produced by the app.
- NOT a data warehouse: uploaded datasets stay in browser memory for the session
  and are not persisted by the app.
- NOT a regression or time-series forecasting platform: the studio targets
  binary/multiclass classification.

## Pipeline steps

1. Upload and profile the dataset (row counts, missingness, distributions,
   impossible clinical values, identifier detection).
2. Clean: imputation (mean, median, KNN), outlier handling with capping,
   biological range filters (for example age 18–100, height 120–220 cm),
   manual column exclusion.
3. Anonymise: PII/PHI privacy shield (40+ patterns), pseudonymisation,
   k-anonymity checks.
4. Choose the target; class balance is checked and a warning is raised beyond
   roughly a 3:1 imbalance.
5. Split, then preprocess training data only (no leakage), with class balancing.
6. Configure algorithms and tuning: grid, random or Bayesian search with
   stratified K-fold cross-validation and regularisation.
7. Evaluate: ROC-AUC, PR-AUC, balanced accuracy, sensitivity/specificity,
   confusion matrix, calibration, MCC. A majority-class baseline audit flags
   when headline accuracy is prevalence-driven rather than real skill.
8. Interpret: SHAP and LIME style attributions, partial dependence/ICE, plus a
   plain-language clinician insight per explanation.
9. Audit fairness: subgroup performance and disparate impact ratios, with
   fairness-aware retraining suggested when a ratio falls below 0.8.
10. Deploy and monitor: generated deployment suite, FHIR/HL7 hooks, simulated
    drift monitoring (retraining suggested when PSI exceeds 0.25).
11. Optional: federated learning simulation (FedAvg with privacy budget
    tracking) and clinical text NLP feature extraction.

## Evaluation guidance PhaBeta enforces

- Overall accuracy is treated as unreliable on imbalanced clinical data. If a
  naive majority-class predictor would score similarly (relative error
  reduction under 20%, or an accuracy-to-balanced-accuracy gap over 8 points),
  the app labels the result prevalence-driven and directs the reader to
  sensitivity, PR-AUC, balanced accuracy and MCC instead.
- Sensitivity (recall for the clinically important class) is prioritised over
  raw accuracy.
- Fairness is weighted ahead of accuracy in the overall score (roughly 60/40).

## Limitations

- Structured/tabular data and clinical free text only — no images, waveforms or
  genomics pipelines.
- Simulated metrics: figures illustrate what a comparable real model would
  likely produce; they are not a validated performance claim.
- Datasets live in browser memory for the session and are not stored.
- No regulatory clearance; outputs must be independently validated before any
  clinical use.

## Frequently asked questions

Q: Can PhaBeta analyse medical images or radiology scans?
A: No. PhaBeta works only on structured/tabular data and clinical free text.

Q: Does PhaBeta train a real model?
A: No. Results are simulated from the dataset's real statistical properties, and
PhaBeta exports Python so the user can train the model in their own environment.

Q: Is PhaBeta safe to use for clinical decisions?
A: No. It is a design, teaching and governance tool. It is not a medical device
and must not be used to make decisions about individual patients.

Q: What data protection rules does PhaBeta follow?
A: UK GDPR, the Data Protection Act 2018, PECR for cookies, IDTA provisions for
transfers, and NHS clinical safety standards DCB0129 and DCB0160.

Q: Who is PhaBeta for?
A: Clinicians, biostatisticians, NHS and health-tech teams and researchers who
need a defensible clinical ML pipeline without an in-house ML engineering team.

## Key pages

- https://phabeta.com/ — platform overview
- https://phabeta.com/what-phabeta-does — capabilities, workflow, FAQ, limitations
- https://phabeta.com/faq — Q&A on clinical validation, fairness and deployment workflows
- https://phabeta.com/healthcare — clinical ML positioning and governance
- https://phabeta.com/clinical-ml-classification — guide plus 12-point checklist
- https://phabeta.com/use-cases/cancer-risk — worked cancer-risk walkthrough
- https://phabeta.com/guides/prevalence-driven-accuracy — why high accuracy on imbalanced clinical data is often prevalence, not skill; majority-class baseline maths and imbalance-robust metrics
- https://phabeta.com/datasets — real clinical datasets by condition and format (UCI, PhysioNet, NHANES) with class balance, access route and an eight-step analysis workflow
- https://phabeta.com/docs — documentation index
- https://phabeta.com/blog — articles on healthcare ML, fairness and deployment
- https://phabeta.com/privacy — UK GDPR and information governance
- https://phabeta.com/contact — support and partnerships