IB World Studies SL Extended Essay Example

The student's question is relevant, but “accuracy and reliability” are not defined operationally. They could specify which metric(s) will determine success, and whether they mean malignant-class recall, macro F1, or weighted performance. That would make the scope more focused and prevent later claims from drifting.

A strong title: the student clearly identifies the topic, the model families, and the application area. This helps the reader immediately understand the essay’s scope and makes the investigation look purposeful from the start.

The methodology is too broad for the research question as written. The student describes a comparative study of secondary sources and their own experiments, but they do not fully explain how the models, datasets, and metrics were selected in a way that ensures fair comparison. They should clarify the decision rules for inclusion and comparison.

If the student keeps the dermatologist comparison, they should make it a consistent analytical thread rather than a passing claim. They could define exactly which published dermatologist benchmark they are using and explain how their model outputs are compared with it.

This is a clear methodological choice: the student states the investigation design directly and explains why it is feasible for an extended essay. That transparency helps the reader follow how evidence will be gathered and interpreted.

This description is only partially accurate because logistic regression is not limited to binary classification in all contexts. The student should avoid oversimplifying the model’s capabilities; instead, they should explain that in their investigation they are using it in a binary setting, which is a methodological choice rather than a universal limitation.

This is a conceptual overstatement. LDA is not automatically better than logistic regression, and the comparison depends on assumptions, feature space, and data distribution. The student should reframe this as a conditional claim supported by evidence from their own results.

This is a strong framing sentence because it shows awareness of why model evaluation in medicine cannot rely on a single headline metric. The student correctly motivates the need for recall and F1, which is exactly the kind of reasoning that strengthens an investigation.

To improve clarity, the student should explicitly state which class is treated as the positive class throughout the essay and keep that convention consistent in every table and discussion. That will make their metrics easier to interpret and compare.

This shows good subject knowledge because the student explains the dataset composition and connects lesion classes to the benign/malignant distinction. That makes later comparisons more meaningful and shows they understand why dataset structure matters in clinical AI.

The explanation is broadly correct, but the wording is imprecise. The student should use standard terminology such as separating classes by maximizing between-class variance relative to within-class variance. Precise language matters here because small conceptual errors weaken the credibility of their technical explanation.

The student shows good understanding of why the assumption is limiting, but the judgement is a little absolute. It would be stronger to explain that Naive Bayes can still work surprisingly well despite violated independence assumptions, especially as a baseline model.

This is a clear and relevant explanation of residual learning. The student connects the architecture to the vanishing-gradient problem, which shows they are not just naming techniques but explaining why the architecture matters.

The formula is not presented in a readable or mathematically standard way. The student should rewrite it clearly as $\frac{TP+TN}{TP+TN+FP+FN}$ and define the symbols once. Clear notation is important because the rest of their metric discussion depends on it.

This table is useful, but the metric labels are incomplete without stating whether the values are macro, weighted, or class-specific averages. In a medical dataset with imbalance, that distinction is essential for understanding what the numbers actually mean.

This is a strong knowledge-based move because the student compares their own outputs with published findings rather than treating their run as isolated evidence. That comparison shows engagement with the wider research base.

This conclusion is not supported by the evidence the student provides later in the appendix, especially for malignant-class recall. They need to distinguish overall or weighted metrics from the clinically important class-specific results, otherwise the knowledge shown here becomes misleading.

This is effective analysis because the student compares their result against a published benchmark and offers a plausible reason for the difference. That combination of evidence and explanation is what turns description into analysis.

This interpretation is not secure unless the student specifies whether these are weighted or malignant-class values. If they are weighted averages, they can hide weak cancer detection performance. The student needs to analyse the class-specific report before claiming the model was strong.

The table is analytically incomplete because it does not state the averaging method for the metrics. That omission matters here: without knowing whether the figures are weighted, macro, or malignant-class values, the comparison can be seriously misleading.

This is a promising analytical observation, but it needs to be tied to the actual confusion-matrix evidence. Explain how many malignant cases were missed and whether the reported F1 refers to the malignant class or a weighted average; otherwise the inference remains too vague.

Good insight, but the sentence stays general. Strengthen the analysis by saying exactly what changes across datasets seem to affect performance here — for example, resolution, class balance, preprocessing, or label distribution — and connect that back to the model outputs.

This is the central claim of the student's essay, but it is not yet justified by the evidence. The appendix shows a model with near-zero malignant recall, so they cannot call the diagnosis reliable without reconciling that contradiction and prioritising the clinically relevant metric.

What changes if the student separates overall accuracy from malignant-class recall? If a model performs well on average but fails to detect melanoma, does that still count as “outperforming” in the context of diagnosis?

This is a useful analytical limitation because the student identifies concrete sources of variance rather than simply saying the study is imperfect. That kind of specificity helps explain why comparisons across studies are difficult.

This is a strong evaluation point because the student identifies a genuine limitation and explains why it matters clinically. Interpretability is not just a theoretical concern here; it affects trust, accountability, and the ability to validate a decision.

This is a balanced evaluation because the student acknowledges a limitation while also drawing a methodological lesson from it. That dual perspective strengthens the credibility of their reflection on the investigation.

This sentence overreaches because the appendix output does not support the claim that CNNs enhance reliability in every case. Evaluate the results more cautiously by separating where the models did well from where they failed, especially on malignant recall.

This comparison is useful, but it needs more careful evaluation. A single benchmark study does not justify a general claim of clinical equivalence, especially when the student's own outputs include contradictory malignant-class performance. They should explain the limits of transferring that evidence to their dataset.

The student correctly identifies class imbalance, but they do not fully evaluate its consequence for their results. They should tie this limitation directly to malignant-class recall and to the risk of inflated weighted averages, so the reader can see how it affects trust in their conclusions.

This conclusion is too strong given the evidence in the student's own appendix. Because malignant-class recall is very weak in at least one model run, they should qualify the claim and explain that performance depends heavily on dataset, metric choice, and evaluation setup.

The student meets the word-count requirement, which shows careful attention to the basic submission conditions.

The title is clearly stated at the start of the document, which is exactly what the format requires.

This second conclusion section repeats material already covered earlier. Remove the duplication and integrate the final synthesis into one clearly labeled conclusion so the structure supports, rather than distracts from, the argument.

The section numbering skips in a way that suggests missing sections. Even though the content is understandable, the structure should be consistent so readers can follow the progression of ideas without confusion.

The contents page and later sections show duplicated and inconsistent section numbering, which makes the document harder to navigate. Clean up the hierarchy so each section appears once and the numbering progresses logically.

Consider adding a note under each results table stating whether the metrics are macro, weighted, or class-specific. That small addition would make the tables much easier to interpret and would prevent serious ambiguity in the discussion.

This output shows a convergence issue in the logistic regression run, but the problem is not addressed in the main body. If a method emits a warning like this, the student should explain whether they increased iterations, scaled features, or otherwise corrected the model before drawing conclusions.

The confusion-matrix output should be discussed directly in the main analysis, not left to speak for itself in the appendix. Since it reveals major class-specific behaviour, it is central evidence and needs to be reconciled with the table values in the essay body.

The confusion matrix and classification report are crucial because they show a different story from the headline table values. The student should bring this evidence into the results section and explain the discrepancy so the reader can trust their presentation of the data.