Collapsed 562 exact duplicate rows while retaining copy counts and grouped identical readings with conflicting labels
VERIFIED CASE STUDY · RESPONSIBLE MACHINE LEARNING
Maternal Health Risk Screening
An honestly tested maternal-risk screening model that removes duplicate leakage, prioritises the cost of missed high-risk mothers and explains how far the result can—and cannot—be trusted.
Pythonscikit-learnpandasTableau
01 · CONTEXT
From a broad concern to testable questions.
Business problem
Can a community health worker use six quickly measured vital signs to identify which mothers need attention first without overstating the model's reliability?
Questions explored
- How much does duplicate leakage inflate apparent model accuracy?
- Which model best balances accuracy with high-risk detection?
- What referral threshold reduces dangerous false negatives?
- Which measurements drive the prediction, and where does the model fail?
Dataset
UCI Maternal Health Risk dataset from rural Bangladesh: 1,014 rows reduced to 452 distinct records after auditing exact duplicates, conflicting labels and implausible readings.
02 · ANALYSIS
A transparent path from preparation to interpretation.
Used repeated grouped cross-validation so identical readings never appeared in both training and test folds
Compared five models and simple clinical baselines, then selected a random forest based on unseen-record performance
Set a 10% referral threshold from the cost of missing a high-risk mother and translated predictions into three triage bands
03 · DASHBOARD
TABLEAU REPORT GALLERY1 decision view
Select any view to inspect the dashboard at full size.
04 · FINDINGS
What the analysis surfaced
01Removing duplicate leakage reduced measured accuracy from 81% to 72%; with duplicates left in, 80% of test rows had an identical copy in the training data.
02At the default 50% threshold the model missed 25 of 112 high-risk mothers; lowering the screening threshold to 10% reduced misses to four while flagging 43% of mothers for review.
03Blood sugar and blood pressure produced 71% of the full model's 72% accuracy, while age, temperature and heart rate added only two points.
04Mid-risk classification remained weak at 24% because the six available vital signs overlap heavily between low- and mid-risk records.
Recommended next moves
01Use the model only as a screening aid to support review, never as a diagnosis or autonomous care decision.
02Use the 10% referral threshold when the priority is reducing missed high-risk mothers, with clinical review for every flagged case.
03Prioritise reliable glucose and blood-pressure measurement because those readings carry most of the predictive signal.
04Improve data capture and collect pregnancy-specific variables—such as gestational age, history and urine protein—before attempting real-world validation.
LIMITATIONS
What the analysis does not establish.
- This is a portfolio project, not a medical device; it has not been clinically validated and must not be used for patient-care decisions.
- The analysis uses 452 distinct records from one region, so results may not generalise to other populations or clinical settings.
- The source does not explain how labels were assigned, and 35 sets of identical readings carry conflicting risk labels.
- Patient IDs, dates, gestational age, medical history and several clinically important measurements are unavailable.
NEXT CASE STUDY
Continue exploring