BMS5010 Artificial Intelligence for Health Science Research and Management

How BMS5010 is assessed

The assessment structure published in the Week 1 lecture for this offering lists these seven components and prints its own total of one hundred per cent, with a separate line recording that the end-of-semester examination carries zero. Every row matches a live submission point on the course site. The word hurdle does not appear anywhere in the course materials, and a numeric attendance threshold is not published there either, although attendance in person at a scheduled demonstration is a stated condition of scoring at all for that component; check the course site for whether either bears on whether you pass. Note also that the syllabus page on the course site displays an older and incompatible table, described in the questions below.

Current BMS5010 dates

Dates are as published in the published course schedule and the submission points on the course site for this offering. Confirm exact deadlines and submission settings in the live LMS.

What BMS5010 covers

Seven chapters follow the published teaching sequence, from what a learning algorithm does through the supervised and unsupervised task types to the error measures for each, generalisation, and the reproducible notebook you demonstrate in person.

The lecture hour covers the concepts of the algorithms and their applications; the laboratory hour is a practical guide to coding, to the individual assignment and to the group project. The course materials are unusually direct about who should not take it, listing prior programming experience as an assumption and heavy independent study outside class as a requirement.

Three quarters of the mark is earned through work you build in Python and then explain in person to an assessor who can ask you anything about it.

How Machines Learn from Health Science Data

The opening chapter separates three words that are used interchangeably outside this room and are not interchangeable inside it.

Artificial intelligence is the engineering ambition, machine learning is the route that learns behaviour from accumulated data rather than from written rules, and deep learning is the branch in which the system learns its own representation of the input instead of being handed features by a person.

The training loop is stated in three steps, see a known answer, measure the gap, adjust the weights, and every method later in the course is a variation on it.

The chapter closes on the five limits the course publishes, of which the two that reach print most often are that these systems find correlation rather than causation and that they are only as good as the data they learned from.

Supervised Learning: Classification and Regression

Supervised learning uses labelled data, and the chapter spends its time on the two questions that fix the task before any library is opened: whether every row carries a label, and whether the thing being predicted is a category or a quantity.

Binary and multiclass problems are separated with the course's own clinical examples, from immune therapy response through tumour appearance to cancer stage, and the continuous side covers inpatient mortality, emergency department demand and survival.

The chapter then works through the four linear models the individual assignment compares, and explains what each penalty on the coefficients actually buys, because a question about regularisation in this subject is usually a question about the error measure in disguise.

Unsupervised Learning: Clustering and Structure

Without labels there is nothing to be wrong against, and that single fact changes the burden of proof.

The chapter builds the two similarity measures the course names, straight-line distance and the angle between feature vectors, and shows that choosing between them is a clinical judgement rather than a coding detail because one keeps magnitude and the other discards it.

It treats the cluster identifier as an output with no explanation attached, so that naming a cluster the severe group becomes a claim requiring evidence from outside the clustering.

Dimension reduction is motivated by the width of biological data, and the chapter is blunt that reducing thousands of measurements to two coordinates makes a picture possible without making the picture a finding.

Measuring Error in a Regression Model

Three measures summarise the same column of mistakes and they can rank two models differently, which makes choosing between them a modelling decision rather than an administrative one.

Average absolute difference treats a miss of six as six times worse than a miss of one; average squared difference treats it as thirty-six times worse; the root of the second puts the figure back into the units of the outcome, which is the only reason it exists.

The chapter carries a worked case in which two models have identical absolute error and very different squared error, and it explains the random-input comparison the assignment asks for, which is the cheapest way to find out whether a model has learned anything at all.

Reading a Confusion Matrix Without Being Fooled

Every metric for a binary decision is one of four counts divided by another, so the chapter teaches the denominators as positions in a table rather than as definitions to memorise.

Ratios that divide by a row total are properties of the test and do not move when the condition becomes rarer; ratios that divide by a column total describe what a result means to the person holding it and move a great deal.

That distinction is worked all the way to the end on a screening cohort, where a test that detects almost everyone and clears almost everyone still produces a positive result that is wrong far more often than it is right, purely because the condition is uncommon.

The chapter finishes on the threshold sweep and the rule for comparing two curves that cross.

Overfitting, Validation and Generalisation

Underfitting, good fit and overfitting are statements about the gap between two measurements, so a model examined on one dataset cannot be placed in any of the three.

The chapter shows the same triple in regression, in classification and in clustering, where the number of clusters turns out to be the same kind of knob as the degree of a polynomial. It then covers the three portions of data and the different job each one does, early stopping at the turn in the validation curve, and the failure that makes everything else look fine, which is information crossing the boundary before the split.

The closing section separates validating inside your own dataset from validating on a cohort from somewhere else, and is clear that only the second answers the transfer question.

Building a Notebook You Can Defend

Thirty of the hundred marks are given for standing in front of somebody and explaining work you are not allowed to edit while you do it, which is more than the mid-term carries.

This chapter treats the notebook as the artefact that has to survive that conversation: a named environment created once and used for everything, the package set the course names, and the structure the assignment specifies of a markdown section, code cells and results for each task.

It lays out the four published ways marks are lost, none of which is about the quality of the analysis, and it closes on the group project, its published topics and named open datasets, the reproducible archive, the report limits and the peer contribution statement.

How to use this guide

Read the front matter first, then work one chapter at a time.

Every chapter ends in practice items with full answers; write your own answer before reading them, because recognising a good answer and producing one under a clock are different abilities and only the second is assessed here.

The practice chapter plans a complete project on an original problem, decision by decision, and the closing chapter is a revision pass for the week that carries the remaining marks.

Pass rules, attendance and what is not published

Three things are worth settling before you plan anything.

The word hurdle does not appear anywhere in the course materials, so no component is published as one; that is a statement about what the materials say rather than a guarantee, and the course site is where to confirm whether any component has to be passed separately.

A numeric attendance threshold is not published anywhere in the course materials either, yet attendance is still a condition: the assessment briefs state that a student who does not attend a scheduled demonstration in person with a student card receives nothing for that component, and that all students must attend the group demonstrations. Check the course site for whether attendance bears on whether you pass overall.

Two further rules carry a zero rather than a deduction, and neither is a sliding scale: work submitted after the published time scores nothing for that component, and a group report whose similarity or generated-content ratios sit above the published thresholds scores nothing for the report and is referred onward.

Evidence and assessment control

Assessment labels and weights follow the assessment structure published for this offering in the Week 1 lecture, which totals one hundred per cent and matches every live submission point on the course site.

The syllabus page on the course site shows a different and incompatible set of components, described in the assessment note below; confirm the current position on the course site before planning around either version. Teaching explanations, worked examples and practice items in this guide are independently authored and are not University assessment material; any marks shown on them are an AskSia study allocation.

The published teaching sequence also includes sessions on multi-omics applications, hands-on work, large language models in medicine and public health, and no teaching material for those sessions was available when this guide was written, so they are not reconstructed here.

Worked example · free

Say what a positive screening result is worth to the person holding it

Key terms

Supervised Learning

Learning from a dataset in which every example carries the correct answer, so the model can be corrected against a recorded outcome during training and scored against one afterwards.

Confusion Matrix

A two by two table of a binary classifier's results, separating correct and incorrect predictions of each class so the trade between false alarms and misses is visible rather than averaged away.

Sensitivity

The share of patients who genuinely have the condition that the model flags, also called recall or the true positive rate. Its denominator is a property of the patients, so it does not move when the condition becomes rarer.

Specificity

The share of patients without the condition that the model correctly clears, also called the true negative rate. Like sensitivity it is computed inside a group defined by the patient rather than by the model.

Precision

The share of flagged cases that were genuine, also called the positive predictive value. Its denominator is everyone the model called positive, so it falls sharply when the condition is uncommon.

Negative Predictive Value

The share of cleared cases that were genuinely clear. It carries most of the useful information in screening, where nearly every result is negative and the guarantee that matters is that a negative really is negative.

Prevalence

The proportion of the tested group that actually has the condition. It leaves sensitivity and specificity untouched and rewrites both predictive values, which is why a model validated in a specialist clinic disappoints in general practice.

Root Mean Squared Error

The square root of the average squared difference between prediction and observation, which restores the units of the outcome and so is the only one of the three regression measures a reader can interpret directly.

Regularisation

Adding a penalty on the size of a model's coefficients to the quantity being minimised. Penalising absolute values drives some coefficients to zero; penalising squares shrinks them all without eliminating any.

Overfitting

Capturing the noise in a training set rather than the pattern beneath it, recognised by strong training performance alongside weak performance on data the model has not seen.

Data Leakage

Information from the held-back data reaching the model before it is scored, usually through a preparation step such as standardising or selecting variables on the whole dataset before splitting it.

External Validation

Testing a model on an independent dataset collected at another hospital, from another population, or in another period. It is the only evidence that a model transfers, and no procedure run inside one dataset can substitute for it.

Silhouette Score

A clustering measure combining how close a point sits to its own group with how far it sits from the nearest other group, running from minus one to one, where values near zero mean the groups overlap.

BMS5010 FAQ

Questions students ask in each BMS5010 chapter

Is deep learning always better than the older approach?

No, and the course frames the difference as a trade rather than an upgrade. Learning the features instead of receiving them needs far more examples, far more computation, and produces a representation nobody can inspect. That price is worth paying when the input is an image, a waveform or a sequence, because nobody can write down what a good feature would be. It is usually not worth paying on a clinical table of a few thousand rows with twenty measurements that clinicians already defined, which is exactly the shape of the individual assignment. Chapter 01 · How Machines Learn from Health Science Data

How do I decide whether an ordered outcome is a category or a quantity?

Ask whether the true underlying values are recoverable. Categories such as no event, one event and more than one are ordered, so treating them as a quantity keeps the information that two is further from zero than one is. But the top category has no upper bound, so any number a regression produces above one is an invention. Predict the count only when the true counts exist in the data; otherwise keep the ordered categories and say in the methods that the ordering is not being used, which is honest and costs you nothing. Chapter 02 · Supervised Learning: Classification and Regression

If clustering cannot be wrong, how is it ever evidence?

By being stable and by predicting something it was not given. Stability means the same grouping survives resampling, a different random start and a modest change to the feature set; a grouping that moves under any of those is a property of the run rather than of the patients. Prediction means the groups differ in a variable that was deliberately held out of the clustering, such as an outcome or a treatment response. Neither check makes clustering a supervised method, and together they are the difference between a hypothesis and a finding. Chapter 03 · Unsupervised Learning: Clustering and Structure

Should I fit on one measure and report another?

Often yes, and it is worth saying so explicitly. Fit on the measure that matches the cost of being wrong, because that is what the optimisation will chase. Report the figure a reader can interpret, which usually means the root of squared error in the units of the outcome. Reporting the absolute error alongside costs one line and answers a different question, namely how the model does on a typical patient rather than on its worst ones. What is not defensible is fitting on one measure and reporting only another without saying which was which. Chapter 04 · Measuring Error in a Regression Model

Which single number should I put in the abstract of a report?

None, if you can avoid it, and if you cannot then the pair that matches the decision. An abstract quoting one figure invites the reader to assume it is the relevant one, and for a rare condition that assumption is usually wrong. The defensible short form names the operating point and gives two numbers: at a threshold flagging one patient in six, the model reached this precision and this recall. That is barely longer than a single figure and it cannot be misread, which is the whole purpose of an abstract sentence. Chapter 05 · Reading a Confusion Matrix Without Being Fooled

Why report the spread across folds and not just the average?

Because the spread is what tells a reader whether the average means anything. Ten folds that agree closely say the estimate is stable and would survive a different random split; ten folds ranging widely say the dataset is too small or too varied for the headline figure to be relied on, and that is a finding worth publishing rather than hiding. The spread also protects you in the demonstration, because it is the honest answer to a question about how confident you are, and it is far stronger than an opinion. Chapter 06 · Overfitting, Validation and Generalisation

What should I actually say in the first thirty seconds of a demonstration?

Name the task, the data and the decision, then say what your result is. Something like: this task predicts a continuous outcome from six routine measurements, I compared four linear estimators on one split, and the penalised fit generalised best by this margin. That sentence orients the assessor immediately and buys you the remaining four minutes to show rather than explain. Starting instead by scrolling through imports spends the most valuable part of a five-minute slot on the least valuable part of the notebook. Chapter 07 · Building a Notebook You Can Defend

How to prepare for the assessments

Carry one small clinical table through all seven chapters and answer each chapter's question about it before reading the chapter's answer: is the outcome a category or a quantity, would grouping tell you anything the label does not, which error measure would a clinician actually care about, and how would you know the model had memorised rather than learned. Then practise the move the assessment is built on.

Open your notebook, pick any cell, and say in one sentence why it is there and what its output means. Doing that aloud for twenty minutes is worth more than another pass over the slides, because it is the only exercise that rehearses what the demonstration marks are given for.

Your AI Artificial Intelligence tutor for BMS5010

Stuck on a hard BMS5010 question? Sia is AskSia’s AI Artificial Intelligence tutor — ask any BMS5010 Artificial Intelligence for Health Science Research and Management 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.