Selected work · Clinical AI + Data Systems

Four projects. Real questions. Honest results.

I am Paul Sentongo, a data scientist and applied AI researcher. I build systems, test what they can actually do, and explain where the evidence is still weak. This portfolio focuses on the four projects that best show how I think.

Read the project stories View my GitHub

Selected research projects

Depth matters more than a long project list.

These projects cover clinical fairness, medical imaging, reproducible machine learning, and trustworthy data infrastructure. The results are presented without hiding failed assumptions or turning prototypes into claims they cannot support.

01 Clinical AI · Fairness · Uganda

Hypertension risk research in a real, incomplete clinical cohort

A study of whether a screening model can remain useful, understandable, and fair when the health records contain serious missing data.

3,471patient records in the Nakaseke cohort

49.3%of age values missing at intake

0.649AUROC on the held-out test group

26.2%overall recall at the tested threshold

The research question

Can a hypertension risk model trained on community health records be trusted when almost half of the age entries are missing—and does it work equally well for different patient groups?

What I did

Compared simple missing-value filling with MICE and a MissForest-style method that learn from relationships between fields.
Kept imputation inside the training pipeline so test patients could not leak information into the model.
Audited performance separately by sex and age instead of trusting one overall score.
Used SHAP to study which features and feature pairs influenced predictions.

The result, in everyday language

An AUROC of 0.649 means the model found some real signal, but it was only moderately better than chance. More importantly, the fairness audit found that it identified just 14% of hypertensive men, compared with 36.2% of hypertensive women, and missed every hypertensive patient under 30 in that test group. That is not a result to hide—it is the strongest reason this model should not yet be used on its own.

Limitations

The data comes from people who reached one hospital-linked screening programme, not a representative sample of rural Uganda.
The study is a single snapshot in time and lacks several useful biomarkers and long-term follow-up.
Age-group findings use only records with observed age, because grouping people by an imputed age would introduce another uncertainty.
The fairness audit currently covers sex and age, but not geography or other access-related groups.

Research opportunities

Collect prospective data across more hospitals and community settings.
Recalibrate the decision threshold and test whether group-aware rules improve sensitivity without creating new harms.
Add geographic access, stronger clinical measures, and repeated patient visits.
Explore retinal imaging only after paired image-and-record data is collected, followed by a new fairness audit.

02 Data Engineering · SQL Server

Turning disconnected CRM and ERP files into a trusted data warehouse

A complete T-SQL pipeline that preserves raw data, fixes inconsistencies, tests quality, and gives analysts a clean model for customer and product questions.

6raw CRM and ERP source files

3Bronze, Silver, and Gold layers

1sales fact table at order-line level

2customer and product dimensions

The research and engineering question

How can data from two systems that use different keys, codes, and formats be turned into one reliable source for reporting—without losing the original evidence or silently accepting bad records?

What I did

Loaded six source files unchanged into a Bronze layer to preserve an audit trail.
Built a transactional Silver process that cleans dates, removes duplicates, standardises codes, and recalculates inconsistent sales values.
Modelled the Gold layer as a star schema with sales, customers, and products.
Added referential-integrity tests, a data catalogue, design documentation, and reusable customer and product reports.

The result, in everyday language

The original files disagree and require cleaning every time someone wants an answer. The warehouse does that difficult work once, records the rules, and presents simple tables for analysis. A user can study sales trends, valuable customers, or product performance without first becoming an expert in the messy source systems.

Limitations

This is a portfolio-scale, batch-loaded warehouse rather than a live high-volume enterprise system.
Source paths must currently be configured for the SQL Server host.
The pipeline does not yet demonstrate change-data capture, automated orchestration, or slowly changing dimensions.
Performance has not been benchmarked at production-scale data volumes.

Research and engineering opportunities

Measure data-quality improvement at each layer and publish reproducible benchmarks.
Add incremental loading, Type 2 history, orchestration, alerting, and lineage.
Test the model against larger workloads and compare query/index strategies.
Connect the Gold layer to a governed BI dashboard and study how definitions affect decisions.

03 Medical Imaging · MLOps

Kidney CT classification built as a reproducible ML system

A VGG16 transfer-learning project that classifies kidney CT images as Normal or Tumour and makes the full training-to-deployment path traceable.

89.9%accuracy on the project validation split

2classes: Normal and Tumour

4versioned machine-learning stages

30%of the dataset used for validation

The research question

Can transfer learning recognise tumour-related patterns in kidney CT images, and can the experiment be packaged so another person can reproduce, inspect, and deploy the same model?

What I did

Adapted a VGG16 model pre-trained on ImageNet to classify kidney scans.
Separated ingestion, model preparation, training, and evaluation into a DVC pipeline.
Tracked parameters, metrics, and model versions with MLflow and DagsHub.
Packaged the application with Flask and Docker and automated deployment to Hugging Face Spaces.

The result, in everyday language

The model classified about nine out of every ten images correctly in its validation split. That is a promising engineering result, but it does not mean it will be 89.9% accurate in a different hospital. Images from a new scanner, patient population, or clinical workflow may look different, so external testing is essential.

Limitations

The model only separates Normal from Tumour; it does not identify cysts, stones, tumour type, or disease stage.
The reported number is validation accuracy from the project dataset, not independent clinical validation.
Accuracy alone does not show how many tumours were missed or how certain the model was.
The current interface does not explain which image region influenced a prediction.

Research opportunities

Evaluate sensitivity, specificity, confidence intervals, and performance by scanner or site.
Validate on an external hospital dataset before making clinical claims.
Extend the classes to cysts and stones and add Grad-CAM explanations reviewed by radiologists.
Monitor image drift and prediction quality after deployment.

04 Dental Imaging · Object Detection

Dental OPG analysis that shows where a suspected condition appears

A YOLOv8 system for panoramic dental X-rays that validates the upload, draws localised findings, and creates a report a clinician can review.

~260annotated OPG images in the detection subset

6dental condition categories

5stages in the DVC pipeline

75/15/10training, validation, and test split

The research question

Can a lightweight detector find and locate several conditions in a panoramic dental X-ray, while rejecting unsuitable images instead of producing a confident-looking but meaningless answer?

What I did

Trained a YOLOv8s detector for caries, fractures, impacted teeth, infection, badly decayed crowns or roots, and healthy teeth.
Added checks for panoramic shape, grayscale appearance, brightness, and minimum image size.
Built colour-coded boxes, plain-language descriptions, recommended actions, and downloadable PDF reports.
Organised ingestion, validation, transformation, training, and evaluation as a reproducible DVC pipeline with MLflow tracking.

The result, in everyday language

The prototype does more than say “there may be a problem.” It marks the area that needs attention and produces a report for review. The repository does not yet publish a final clinically validated accuracy score, so the honest conclusion is that the workflow is working but its diagnostic reliability still needs to be established.

Limitations

About 260 annotated images is small for representing the full variety seen in dental practice.
There is not yet a reported external test on images from a different clinic or scanner.
Per-condition precision, recall, and detection performance still need to be published clearly.
The tool is an aid for qualified professionals, not a replacement for examination and diagnosis.

Research opportunities

Build a larger, diverse OPG dataset with annotations from multiple dental experts.
Measure agreement between annotators and publish per-condition confidence intervals.
Test the model across clinics, machines, and image-quality conditions.
Add uncertainty warnings and study whether the tool improves reading time or consistency in real workflows.

How I approach research

A useful result is more than a high score.

My strongest work connects careful engineering with honest interpretation. I want to know not only whether a model runs, but who it works for, what could break it, and what evidence is still needed.

Start with the real decision

Define who will use the system, what they need to decide, and what a harmful error would look like.

Make the work reproducible

Version data and models, record assumptions, and keep training and evaluation steps traceable.

Translate metrics into meaning

Explain what a number says in practical terms—and what it does not prove.

Treat limitations as findings

A subgroup failure, weak external validity, or missing measure is a research direction, not something to bury.

Research and collaboration

Let’s work on a question that matters.

I am interested in clinical AI, medical imaging, responsible machine learning, MLOps, and data systems for settings where resources and high-quality data are limited.

Email me LinkedIn GitHub