BME5150 Medical Artificial Intelligence and Data Analytics
The Complete Study & Assessment Guide · Semester 1, 2026/27BME5150 Overview
- The Hong Kong Polytechnic University
- Semester 1, 2026/27
- Postgraduate course
- Computer Science
- Project centred assessment
What the course asks you to connect
Medical Artificial Intelligence and Data Analytics joins machine-learning technique to a clinical evidence chain. The teaching materials move from medical images and other patient data through preparation, feature engineering, modelling, evaluation and decision support.
- Assessment Three individual in-class quizzes carry 30%, the project report and code carry 50%, and participation carries 20%.
- Core habit Keep every transformation inside development data and reserve untouched cases for the final evaluation.
- Project standard Choose the best-performing model by a clinical criterion, then state the population and limitation that bound the claim.
- Revision route Recompute small metric examples by hand before interpreting larger model tables.
How BME5150 is assessed
| Component | Weight | Format |
|---|---|---|
| Three In-class Quizzes | 30% | Individual; weeks 4, 8 and 12 |
| Project Report + Code | 50% | Report and code together; week 13 |
| Class Participation | 20% | Attendance, in-class engagement and discussion |
The three published numeric rows total 100%. The project guidance gives the report format, late and similarity rules; use the current course site for submission settings.
Current BME5150 dates
| Date | Item | Control |
|---|---|---|
| Week 4 (23 September) | In-class quiz | Day and month printed in the assessment slide. |
| Week 8 (21 October) | In-class quiz | Day and month printed in the assessment slide. |
| Week 12 (18 November) | In-class quiz | Day and month printed in the assessment slide. |
| Week 13 (25 November) | Project Report + Code | Week and day/month printed in the assessment slide and report guidance. |
The captured files print week and day/month values but no explicit year; verify current submission times on the course site.
What BME5150 covers
Seven chapters move from medical imaging and radiomics through data preparation, classifier evaluation, regression, regularisation and decision trees to the project report.
Medical AI Data and Radiomics
medical images, radiomic features, heterogeneity, federated learning and external validationStart with the data-generating process
Medical artificial intelligence is built from measurements created by clinical workflows. Images depend on acquisition hardware, protocol and reconstruction. Labels depend on readers, tests or later outcomes.
- Medical data types and intended use
- Radiomics from images to quantitative features
- Acquisition and segmentation heterogeneity
- Data augmentation and simulation
- Federated learning and privacy boundaries
- Internal and external validation
Preparing Clinical Data for Modelling
train and test boundaries, scaling, feature reduction, imbalance and reproducible preprocessingSplit before learning transformations
Normalisation, standardisation, imputation and feature selection learn from observed data. If they use the full table, held-out cases influence the pipeline before evaluation begins.
- Patient-level train and test separation
- Normalisation and standardisation
- Missing data and outliers
- Supervised and unsupervised filtering
- High-dimensional radiomic features
- Class imbalance and resampling
Evaluating Classifiers and Explanations
confusion matrices, sensitivity, specificity, precision, ROC curves, AUC and SHAPBegin with the four classification counts
True positives, false negatives, false positives and true negatives preserve the errors hidden by accuracy.
- Confusion-matrix counts and denominators
- Sensitivity, specificity and precision
- Accuracy under class imbalance
- ROC curves and AUC
- Threshold selection and clinical workload
- SHAP and feature-importance boundaries
Regression Error, Bias and Variance
residuals, MAE, MSE, RMSE, R squared, underfitting and overfittingResiduals preserve direction
A residual is the observed outcome minus the fitted prediction. Positive and negative values show which side of the prediction line a case falls on.
- Observed and predicted values
- MAE, MSE and RMSE
- Percentage errors and near-zero outcomes
- Correlation and R squared
- Bias and variance patterns
- Residual diagnostics and held-out testing
Ridge, Lasso and Elastic Net
L1 and L2 penalties, coefficient shrinkage, sparse selection and tuningRegularisation changes the fitting objective
Least squares rewards close fit to development outcomes. Regularisation adds a cost for coefficient size. Lambda controls how strongly that cost competes with fit.
- Penalised regression objectives
- Lambda and coefficient shrinkage
- Ridge L2 behaviour
- Lasso L1 sparsity
- Elastic net mixing
- Scaling, tuning and selection stability
Cross Validation and Decision Trees
fold rotation, leakage control, entropy, information gain, Gini impurity and pruningCross-validation rotates a development boundary
In k-fold cross-validation, development data are divided into parts. The pipeline fits on all but one part and validates on the remaining part, rotating until each part has served as validation.
- Five-fold rotation and repeated validation
- Grouped and stratified folds
- Nested preprocessing and tuning
- Recursive decision-tree splits
- Entropy, information gain and Gini
- Gain ratio and pruning
Project Design, Critique and Reporting
problem framing, method justification, model comparison, clinical criticism and concise reportingOne claim should organise the report
The project guidance asks for the importance of the prediction problem, relevant prior models, adopted methods, comprehensive evaluation, justified selection and criticism of clinical limitations.
- Prediction question and clinical importance
- Comparative literature review
- Method modifications and reproducibility
- Comprehensive model evaluation
- Best-performing model justification
- Clinical limitations and future tests
The central skill is not merely naming an algorithm. It is showing how acquisition, annotation, splitting, preprocessing and metric choice shape the claim that can be made about a model.
Why radiomics is a useful spine
Radiomics turns medical images into quantitative descriptors of shape, intensity and texture.
That makes the full workflow visible: images are acquired, relevant anatomy or disease is segmented, features are extracted, redundant or unstable variables are reduced, models are trained and performance is evaluated. Each step can introduce site-specific information or measurement variation.
External validation therefore tests the complete locked pipeline rather than an isolated classifier.
How the mathematical chapters fit
The course covers confusion-matrix metrics, ROC curves and AUC, linear-regression errors, bias and variance, ridge and lasso penalties, elastic net, cross-validation and decision-tree split criteria. These methods answer different questions.
Sensitivity and precision use different denominators. MAE and RMSE punish large residuals differently. Ridge shrinks coefficients smoothly, while lasso can set them to zero.
Entropy and Gini quantify node impurity, but a tree still needs pruning and held-out evaluation.
What the project report must demonstrate
The published guidance asks students to establish the importance of the prediction problem, explain the adopted methods and modifications, evaluate models comprehensively, justify the best-performing model, and criticise quality and clinical limitations.
The report and code are submitted together. A strong submission makes every result reproducible and every limitation testable. It identifies a likely data shift, explains how that shift would affect the pipeline and names the external or prospective evaluation needed next.
Worked example · free
Choose a threshold from clinical consequences
- 2Complete the four confusion-matrix counts.
- 3Compute the three requested metrics with their denominators.
- 3Translate the false negatives and false positives into a clinical trade-off.
Key terms
- Radiomics
- Radiomics is the extraction and analysis of quantitative features from medical images for modelling or clinical investigation.
- External Validation
- External validation evaluates a locked pipeline on data separated from model development by site, time or another meaningful boundary.
- Data Leakage
- Data leakage occurs when information unavailable at intended prediction time or from held-out cases influences model development.
- Sensitivity
- Sensitivity is the proportion of actual positive cases that the model correctly identifies.
- Specificity
- Specificity is the proportion of actual negative cases that the model correctly rejects.
- Precision
- Precision is the proportion of positive predictions that are truly positive.
- Calibration
- Calibration is the agreement between predicted probabilities and observed outcome frequencies.
- Residual
- A residual is the observed outcome minus the value predicted by a regression model.
- Ridge Regression
- Ridge regression adds an L2 penalty that shrinks coefficients, usually without setting them exactly to zero.
- Lasso Regression
- Lasso regression adds an L1 penalty that can shrink selected coefficients exactly to zero.
- Cross Validation
- Cross validation rotates held-out development folds to estimate performance and choose model settings.
- Entropy
- Entropy measures uncertainty in a node and falls as its class distribution becomes purer.
- Pruning
- Pruning removes or prevents weak tree branches to reduce complexity and improve generalisation.
BME5150 FAQ
How is the course assessed?
The published assessment table assigns 30% to three individual in-class quizzes, 50% to the project report and code submitted together, and 20% to class participation through attendance, engagement and discussion. The numeric rows total 100%, and the project guidance supplies additional reporting and submission rules.
What should I revise before an in-class quiz?
Rebuild the relevant workflow from memory, then work a small calculation by hand. For metrics, write the confusion-matrix cells and denominators. For regression, compute residual summaries. For regularisation and trees, explain what the penalty or impurity criterion changes before interpreting a result.
Why is splitting data before preprocessing so important?
Scaling, imputation and feature selection learn from observed values. If held-out cases influence those decisions, the evaluation is no longer independent and performance becomes optimistic. Split first, fit every learned transformation within development data, and apply the locked sequence to validation or test cases.
Which classification metric is best for medical AI?
No single metric is always best. Sensitivity emphasises detected positive cases, specificity rejected negative cases, and precision the reliability of positive alerts. AUC summarises ranking across thresholds but does not choose an operating point. Match the metric to the clinical consequence and report the threshold.
Does lasso identify the biologically important features?
Not by itself. Lasso identifies variables retained by one penalised predictive model under a particular scale, penalty and set of correlated alternatives. A zero coefficient does not prove biological irrelevance. Check selection stability across resamples and use independent evidence for biological interpretation.
What makes a strong project limitation?
A strong limitation names a mechanism, such as scanner shift, annotation disagreement or prevalence change, and then specifies the test that would reveal its effect. Replace generic calls for more data with a defined external, temporal or prospective cohort and the metric that should be repeated.
How should I justify the best-performing model?
Define performance from the clinical task before comparing results. Apply candidate pipelines to the same held-out cases, state the chosen metric and threshold, compare the nearest alternative, and explain whether the improvement justifies extra complexity, instability or loss of interpretability.
Is a high AUC enough to claim clinical usefulness?
No. AUC describes ranking across thresholds. Clinical use still needs an operating point, threshold-specific errors, calibration when probabilities matter, and evidence that the locked pipeline transports to the intended population. A high internal AUC can coexist with poor precision or external failure.
Questions students ask in each BME5150 chapter
- Why can a model learn the hospital instead of the disease?
- Hospital-specific scanners, protocols, reconstruction and patient pathways can correlate with the outcome. A flexible model may use those easier signals. Hold out sites, inspect feature distributions and report site-wise errors to test whether apparent disease prediction depends on institutional artefacts. Chapter 01 · Medical AI Data and Radiomics
- Why must preprocessing be fitted inside cross-validation?
- Validation folds are meant to represent unseen data. If their values affect scaling, imputation or feature selection, the model receives information about them indirectly. Refitting the complete preparation sequence in every training fold gives a more honest development estimate. Chapter 02 · Preparing Clinical Data for Modelling
- Why can sensitivity and precision tell different stories?
- Sensitivity begins with actual positive cases, while precision begins with positive predictions. A model can detect most positive cases and still produce many false alarms, especially when prevalence is low. Always name the denominator and report the underlying counts. Chapter 03 · Evaluating Classifiers and Explanations
- When should I prefer RMSE to MAE?
- Prefer RMSE when large errors deserve extra influence and that weighting matches the clinical consequence. Prefer MAE for a linear, unit-preserving summary that is less dominated by extremes. Reporting both can reveal whether a small number of large misses drive performance. Chapter 04 · Regression Error, Bias and Variance
- Why should predictors be scaled before ridge or lasso?
- Penalty size depends on coefficient magnitude, and coefficient magnitude depends on predictor units. Without deliberate scaling, the model penalises variables differently for unit reasons. Fit scaling within each development fold and apply the stored parameters to held-out data. Chapter 05 · Ridge, Lasso and Elastic Net
- Is cross-validation a replacement for a final test set?
- Not when it is used to choose algorithms, preprocessing or hyperparameters. Cross-validation is then part of development. Preserve an untouched evaluation set, refit the selected pipeline on development data and apply it once for the final estimate. Chapter 06 · Cross Validation and Decision Trees
- How should I organise a short project report?
- Use the evidence chain: clinical question, comparative literature, reproducible method, held-out results, model choice and testable limitation. Let tables compress repeated numbers and use prose for reasoning. Remove background that does not change a method or interpretation. Chapter 07 · Project Design, Critique and Reporting
How to prepare for the assessments
Use a three-pass routine. First, draw the pipeline and label where each quantity is learned: split, scaling, selection, tuning and evaluation. Second, recompute one small example without notes, checking denominators, signs and units before interpretation. Third, write a clinical boundary sentence that states the population, threshold or evaluation rule, and the external check still required.
Rotate the chapters by dependency rather than by file order: radiomics and preparation first, classifier metrics next, regression and regularisation together, then cross-validation and trees. For the project, keep a living claim sheet beside the code. Record the question, outcome definition, split, transformations, candidate models, selection criterion, held-out result and main limitation.
Before submission, run the code from a clean start and confirm that every table and figure can be traced to generated output. Finish by reading the conclusion as a sceptical clinician: ask who the result applies to, what error it tolerates and what change in site, scanner, annotation or prevalence might break it.
Your AI Computer Science tutor for BME5150
Stuck on a hard BME5150 question? Sia is AskSia’s AI Computer Science tutor — ask any BME5150 Medical Artificial Intelligence and Data Analytics question and get a clear, step-by-step explanation grounded in how the course is actually taught and assessed. Read this whole study guide free, then take your hardest questions to Sia.
