A complete data science thesis has seven sections: title and abstract, problem statement, research questions, related work, methodology (data, features, models, evaluation), results, and discussion with limitations. Below is a full annotated walkthrough built around a fictional study predicting hospital readmission risk from structured health-record data, comparing an interpretable model against a black-box one.
A note before you read further. Every dataset size, metric value, and finding below is a placeholder marked [figure]; every citation you would supply is marked [cite]. This is a shape to copy, not a real study — no dataset, institution, or result named here is real.
Step 1: What does the title and abstract look like?
Title: “Interpretable Versus Black-Box Machine Learning for 30-Day Hospital Readmission Prediction: A Comparison on Structured Electronic Health Record Data.”
Abstract (structured, ~250 words): “Background: Hospitals use readmission-risk prediction to target post-discharge follow-up resources, but clinicians are often reluctant to act on predictions they cannot interpret [cite]. Objective: To compare the predictive performance of an interpretable model (logistic regression with regularization) against a black-box model (gradient-boosted trees) for 30-day readmission risk, and to quantify the performance cost, if any, of interpretability. Methods: Structured EHR data (n = [figure] admissions) was used to train both model classes on the same feature set, evaluated by 5-fold cross-validation on AUROC and calibration. Results: The black-box model achieved an AUROC of [figure] versus [figure] for the interpretable model ([statistic, cite]); calibration differed by [figure]. Conclusion: [One sentence tied to the actual performance gap found, not beyond it.] Keywords: readmission prediction, interpretable machine learning, gradient boosting, electronic health records, model calibration.”
Why it works: the abstract states the applied tension driving the whole thesis in its first sentence — clinicians want accuracy but also want to trust the model — before a single method is named, which is what makes a data science abstract read as a contribution rather than a technique demo.
Step 2: What does the problem statement look like?
“Hospital readmission within 30 days of discharge is a widely used quality and cost metric, and machine-learning models can predict individual readmission risk with reasonable accuracy using data already captured in the electronic health record [cite]. However, the models with the strongest published predictive performance are typically gradient-boosted trees or neural networks, whose predictions are difficult for a treating clinician to interpret or contest, while the models clinicians trust and are willing to act on — logistic regression, simple risk scores — typically show weaker discrimination [cite]. This creates a practical dilemma for hospitals deciding which model class to deploy: accuracy or interpretability, without a clear, quantified answer to how much accuracy interpretability actually costs on a comparable feature set. This study addresses that gap directly, on a single dataset with both model classes trained on identical inputs.”
Why it works: the paragraph names a real deployment decision hospitals actually face, states the accuracy-interpretability tension as the specific unresolved thing (not “more research is needed”), and commits the study to answering it on a controlled comparison — same data, same features, only the model class varies.
Step 3: What do the research questions look like?
- What is the difference in discrimination (AUROC) between a gradient-boosted-tree model and a regularized logistic-regression model for 30-day readmission prediction, trained on identical structured EHR features?
- How does calibration — the agreement between predicted risk and observed outcome frequency — differ between the two model classes?
- Which individual features drive the largest share of predictive performance in each model, and do the two model classes agree on which features matter most?
Why it works: each question maps to one results-chapter table (a performance-comparison table, a calibration plot, a feature-importance comparison), so the methodology section that follows can be written directly against these three questions rather than a vague “build and compare models.” For more on how a research question should be built to be directly testable, see worked research-question and hypothesis examples.
Step 4: What does the related-work section look like?
“Prior work on readmission prediction spans two largely separate literatures. Clinical-informatics studies [cite] typically use simple, interpretable risk scores (e.g., the LACE index) validated for clinical usability but with modest discrimination (AUROC typically [figure]–[figure]). Machine-learning-focused studies [cite] report higher discrimination using gradient boosting or deep learning but rarely report calibration, and rarely compare directly against an interpretable baseline trained on the identical feature set — most comparisons instead cite the interpretable model’s performance from a different paper, on different data, which confounds the comparison. This thesis closes that specific methodological gap: a same-data, same-feature, head-to-head comparison with both discrimination and calibration reported for both model classes.”
Why it works: the paragraph identifies the exact methodological flaw in how the existing literature compares model classes — cross-study rather than same-data comparisons — which is a sharper, more defensible gap than simply claiming no one has looked at readmission prediction before.
Step 5: What does the methodology section look like?
“Data: structured EHR data from [figure] admissions at [named or de-identified source], including demographics, comorbidity codes, prior-admission count, discharge disposition, and lab values at discharge. Outcome: readmission to any hospital within 30 days of discharge, defined per [named standard definition, cite]. Feature engineering: comorbidity burden summarized via the Elixhauser index [cite]; lab values aggregated to last-value-before-discharge; missing values imputed using [named method], with missingness itself retained as a feature where informative — the full logic for choosing and reporting a feature this way follows the same table shown in operationalising variables: table and worked examples. Models: (1) L2-regularized logistic regression, hyperparameters tuned via nested cross-validation; (2) gradient-boosted trees (XGBoost), same tuning approach on the identical feature set. Evaluation: 5-fold cross-validation, reporting AUROC with 95% confidence intervals, calibration plots, and Brier score. Interpretability comparison: logistic-regression coefficients versus SHAP values for the gradient-boosted model, compared for feature-ranking agreement using Spearman’s rank correlation. Ethics: the study uses coded, de-identified retrospective data that remain potentially identifiable at the source, so REB approval is sought for secondary use without consent under TCPS 2 Article 5.5A, which requires the researcher to satisfy the REB that, among other conditions, seeking consent is impossible or impracticable and privacy will be safeguarded; had the extract been fully anonymous, Article 2.4 would have removed the need for REB review.”
Why it works: naming the specific imputation method, the specific comorbidity index, and the specific interpretability-comparison statistic (Spearman’s rank correlation between SHAP and coefficient rankings) turns “we compared feature importance” into a reproducible procedure a reader could actually re-run, and the ethics line identifies which TCPS 2 provision applies to secondary use of coded retrospective data — and why the anonymous-data exemption does not — rather than treating consent as automatically required.
Step 6: What does a results excerpt look like?
“The gradient-boosted model achieved an AUROC of [figure] (95% CI [figure]–[figure]) compared with [figure] (95% CI [figure]–[figure]) for logistic regression, a difference of [figure] points ([statistic, cite]). Calibration plots showed [both models well-calibrated / the gradient-boosted model overconfident at high predicted risk], with a Brier score of [figure] versus [figure]. The top five features by SHAP value in the gradient-boosted model were [feature list]; the top five by absolute coefficient in logistic regression were [feature list], with a Spearman rank correlation of [figure] between the two rankings, indicating [strong / moderate / weak] agreement on which features matter most.”
Why it works: reporting the confidence interval alongside every point estimate, not just the point estimate itself, is the detail that separates a data-science results chapter examiners trust from one where every number reads as suspiciously exact — an interval signals the author understands the estimate has uncertainty.
Step 7: What does the discussion and conclusion look like?
“The [figure]-point AUROC gap between the gradient-boosted and logistic-regression models represents the quantified cost of choosing interpretability in this setting, smaller/larger than the gap reported in [prior cross-study comparison, cite], consistent with this study’s argument that cross-study comparisons on different data likely [over/understate] the true gap. The [strong/moderate] rank agreement between SHAP and coefficient-based feature importance suggests the two model classes are learning broadly similar risk signals, which is itself a finding relevant to deployment: if the models agree on what matters, a hospital may reasonably deploy the interpretable model with only a modest accuracy cost. Limitations: single-site data limits generalizability to other hospital systems with different patient populations or documentation practices; retrospective design cannot evaluate real-world clinical impact on actual readmission rates, only predictive accuracy; missingness patterns may not transfer to a different EHR system’s data-entry conventions. Conclusion: for this dataset and feature set, choosing interpretability cost [figure] AUROC points, a gap [small/large] enough that [recommendation tied directly to the finding].”
Why it works: the conclusion states a number, not a vague verdict — “cost X points” is falsifiable and specific, which is exactly what a data science committee wants instead of an unquantified “interpretability is worth the tradeoff.”
What five things do examiners flag in a data science thesis like this one?
| What examiners flag | The fix |
|---|---|
| A model comparison reported without confidence intervals on the metric | Report the interval alongside every point estimate, from cross-validation folds or bootstrapping |
| Feature importance reported for one model class only | Compare rankings across both model classes with a stated agreement statistic |
| Calibration ignored in favour of discrimination (AUROC) alone | Report a calibration plot and Brier score alongside AUROC |
| An imputation method used without justification for why it fits the missingness pattern | State whether data is missing at random and why the chosen method suits that pattern |
| A retrospective study’s findings stated as if they proved real-world clinical impact | State explicitly that predictive accuracy was evaluated, not deployment impact |
Ready to build your own thesis section by section instead of staring at a blank page? Start free with Tesify and work through each chapter with the structure above as your checklist.
Frequently asked questions
Does every data science thesis need to compare two model classes?
No — many data science theses focus on a single model, a new architecture, or a purely applied deployment question; the comparison structure above suits theses specifically about the accuracy-interpretability tradeoff, which is one common but not universal thesis shape in the field.
What is SHAP, and why is it used alongside coefficients?
SHAP (SHapley Additive exPlanations) attributes a black-box model’s prediction to individual input features using a game-theoretic approach; it is compared against logistic-regression coefficients here because both are, in their own way, a ranking of which features matter most, allowing a like-for-like comparison across model classes.
Do I need REB approval to use retrospective, de-identified hospital data?
Usually yes if the data are coded or otherwise potentially identifiable: TCPS 2 Article 5.5A sets the conditions for using identifiable information without consent. Research relying exclusively on secondary use of genuinely anonymous information does not require REB review under Article 2.4 — your REB decides which applies. See the full TCPS 2 requirements guide for what each route requires you to demonstrate.
How is cross-validation different from a single train-test split?
Cross-validation trains and tests the model on multiple different splits of the data and averages the result, which produces a more stable performance estimate and, critically, lets you report a confidence interval — a single train-test split gives you one number with no sense of how much it might vary.
What is calibration, and why does it matter separately from accuracy?
A model can rank patients correctly by risk (good discrimination, high AUROC) while still being systematically over- or under-confident in its actual predicted probabilities (poor calibration); a hospital deciding who gets a follow-up call based on a predicted 70% risk needs that 70% to actually mean 70%, not just a correctly ordered ranking.
Can I use this structure for a non-healthcare data science thesis?
Yes — the section order (title/abstract, problem statement, research questions, related work, methodology, results, discussion) is domain-general; only the specific content (features, models, ethics considerations) changes with the application area.
What is the biggest structural mistake in a first-draft data science thesis?
Reporting a single headline metric (usually accuracy or AUROC) without uncertainty, calibration, or a stated comparison baseline — examiners read a bare metric as evidence the tradeoffs were not considered, not just omitted from the writeup. If your quantitative design more broadly still needs its statistical-test justification written up, the reasoning process in which statistical test should you use transfers directly, even outside education.