The Rounds Journal · Health data

A healthcare data analyst portfolio project using synthetic data

Build one portfolio example that shows how you define a question, check records and explain uncertainty.

Written byThe Rounds editorial team
Published
Reading time5 min read

Build one portfolio example that shows how you define a question, check records and explain uncertainty.

Choose a decision your analysis could support

Create a fictional clinic dataset with appointment ID, week, scheduled time and attendance status. Ask whether missed appointments appear more common in one time band. Define in advance whether the denominator includes cancelled appointments and how you will handle unknown attendance.

Use a spreadsheet or a tool you can explain in an interview. Keep the original synthetic dataset and document how you made it. Do not copy workplace records, screenshots or identifiers into a public portfolio.

Show the checks beside the findings

Check that appointment IDs are unique and status labels are consistent. Count the records you exclude and say why. For each time band, show the raw count and denominator alongside the percentage.

In a fictional comparison, 4 missed appointments out of 10 and 12 out of 40 give 40% and 30%. The smaller group is sensitive to any single record. Show that limitation instead of treating a ten-point difference as proof of a scheduling problem.

Package the work for a hiring manager

Include a one-page summary, the data dictionary and the steps needed to reproduce the result. Use one chart if it makes the comparison easier. Finish with what you would check next, such as whether the mix of appointment types differs across time bands.

Compare the completed project with actual role descriptions. If a job asks for SQL, explain which part you could rebuild in SQL. Keep the project clearly labelled as practice.

Read the health data analytics study guide for more context.

Plan the files before you start

Use a short README that states the question, how you generated the data and how to reproduce the analysis. Keep the synthetic input separate from the cleaned file. Record cleaning decisions such as how you handled unknown attendance, even if the final presentation is just a spreadsheet and one chart.

Add a sensitivity check

Recalculate the rate after changing one reasonable assumption about unknown cases. If the result changes substantially, explain why it matters. You can also split results by appointment type to see whether the groups look comparable, provided you give readers the counts behind each slice.

Rehearse the interview walkthrough

Start with the decision, show one check you made, then explain the result and one limitation. Ask a colleague to reproduce a number from the files. If they cannot find the denominator or filtering rules, revise the README and chart labels before sharing the project.

Make the case brief and data dictionary part of the submission

Start with a one-paragraph brief. For the appointment example, state that a fictional service is considering whether to investigate booking practice in one time band. Record what your data contains, which fields you invented and what you mean by a missed appointment. Specify the denominator, such as attended plus missed appointments with a known outcome. Write these rules before calculating a result so a reviewer can tell what the analysis was meant to answer.

In your data dictionary, give every field a name, type and example. Include the permitted status labels and the time bands you intend to compare. If the scheduled time is missing, state how you will handle the record. Keep a separate copy of the data as first generated. This lets a reviewer see what you changed rather than starting with an already cleaned table.

Build checks a reviewer can rerun

Check that the appointment identifier is unique and investigate any repeated row. Record the number of rows in the original data and after each major step. Check that the time-band groups add up to the records included in the analysis. If they do not, compare the records affected by date parsing or a missing time. Save the steps as a short script, spreadsheet formula notes or a reproducible notebook. Choose a format you can explain confidently.

Suppose you use SQL to group the records. Include the query that calculates counts by status and time band. Show which filters apply before the grouping. If you later join a table of reminder attempts, check whether one appointment can match several attempts. Summarise the attempts at the appointment level before using them in an appointment-level denominator. Record how you verified that the join did not increase the number of appointment rows.

Give the reader a conclusion with boundaries

Put the counts beside any percentage you display. Explain that a comparison of the fictional sample can demonstrate the method and how you dealt with uncertainty. It cannot establish a real clinic’s attendance patterns. If a potential employer asks how you would take the work into a real service, describe the approved data access, definitions and checks you would need to agree before writing the analysis.

Test the portfolio by giving someone the brief, dictionary and results, then asking them to find one numerator and denominator and reproduce the rate. Ask what they would do with the information and what further evidence they want. Make a short revision after the test and keep the note explaining it. You can use that change in an interview as an example of improving a report based on feedback instead of merely listing the software you used.

Explore more resources →