← SHANNONWAREARTICLE · 2026.11.22

Machine Learning Zoomcamp - Midterm Project Peer Review 3

I finish my peer review of Machine Learning Zoomcamp midterm projects with Breast Cancer Malignancy Prediction Service 🎗️ from katnnabeth.

Peer Review Number 3

Personally, I found this to be (circumstantially, the last, but also) the most well organised and well presented project that I scored for the midterm project. And that is saying something because I found all of the projects to be well laid out and satisfactorily complete.

Interestingly, when I asked my cloud agent to verify my findings, it was more critical, and in particular drilled down on the false positive rate: "The deployed threshold contradicts the project's own analysis. The README says a false negative (a missed cancer) is the most serious error. In the notebook's threshold table, 0.3 gives 0 false negatives on the validation set, against 2 at 0.5. Yet model.bin saves optimal_threshold: 0.50, and the test set then shows 4 missed cancers. The developer should either deploy the lower threshold or explain why 0.5 was kept."

However, there are are other issues that will be born out by re-running the project code:

  1. Pipfile.lock doesn't exist. The README's file listing and its dependency-management row both mention it.
  2. The dependencies aren't pinned. The README says "pinned versions", but requirements.txt and Pipfile only set minimums (>=). model.bin is a pickle created with pandas 2.2.0 and NumPy 1.26.4, so a fresh install may get newer versions that warn when loading it or fail to load it. Exact versions, or the missing lock file, would fix this.
  3. There is no LICENSE file, so the MIT badge links to a page that doesn't exist.
  4. The clone instructions use a placeholder (<your-username>/breast-cancer-prediction) instead of the real repository URL. The Render steps start with git init and pushing the repo, which applies to the author rather than a reviewer.
  5. "Clinical calibration" and "zero overfitting" overstate the work. The notebook never checks calibration, and nothing shows the model can't overfit. Both should be softened or backed up, for example with a calibration curve.
$ jupyter nbconvert --to notebook --execute midterm-project-notebook.ipynb --output /tmp/out.ipynb
...
ModuleNotFoundError: No module named 'matplotlib'
$ pip install matplotlib
...
$ pip install seaborn
...
Successfully installed seaborn-0.13.2
...
$ jupyter nbconvert --to notebook --execute midterm-project-notebook.ipynb --output /tmp/out.ipynb
[NbConvertApp] Converting notebook notebook.ipynb to notebook
/home/codespace/.local/lib/python3.14/site-packages/nbformat/validator.py:434: MissingIDFieldWarning: Cell is missing an id field, this will become a hard error in future nbformat versions. You may want to use `normalize()` on your notebooks before validations (available since nbformat 5.1.4). Previous versions of nbformat are fixing this issue transparently, and will stop doing so in the future.
  _validate(nbdict, ref, version, version_minor, relax_add_props)
[IPKernelApp] WARNING | Kernel is running over TCP without encryption. All communication (including code and outputs) is sent in plain text and is susceptible to eavesdropping. Use IPC transport or launch with kernel manager-provisioned CurveZMQ keys to enable transport encryption.
[NbConvertApp] Writing 49959 bytes to /tmp/out.ipynb
$ python train.py --output-dir artifacts
Loading data from /workspaces/machine-learning-zoomcamp-2026-midterm-project/data/breast_cancer.csv...
Dataset shape: 569 rows, 32 columns
Total features: 30
Target distribution: Malignant = 212 (37.3%), Benign = 357 (62.7%)

Training set: 455 samples | Test set: 114 samples

Training final Logistic Regression model (C=0.1, solver='lbfgs')...

==================================================
FINAL TEST EVALUATION METRICS:
==================================================
  Test ROC-AUC:   0.9977
  Test Accuracy:  0.9825 (98.25%)
  Test Precision: 1.0000
  Test Recall:    0.9524
  Test F1-Score:  0.9756

Confusion Matrix:
  TN=72  FP=0
  FN=2  TP=40
==================================================

Model and scaler successfully exported to /workspaces/machine-learning-zoomcamp-2026-midterm-project/model.bin

The above extract from the bash session has me satisfied with respect to Reproducibility. Only a few dependencies were not in the Codespace as loaded.

Final word: One of the things I learned from this project is that the badges are generated with the use of shields.io. This makes a great first impression when reading the README.

ImpressumContactTermsPrivacy