Dr. D.Y. Patil Vidyapeeth Students: Cleaning Healthcare Data Without Breaking Privacy

You can clean healthcare data safely if you first confirm you are allowed to use it, remove or transform every direct identifier, test for indirect re-identification, and write down what you did. In a short internship, the safest route is usually a synthetic or already-anonymised dataset, plus a signed data-handling note that records your choices.

This guide is for students of Dr. D.Y. Patil Vidyapeeth and others who want to practise health analytics. It gives a de-identification checklist, a consent and ethics question list, a sample note you can sign, and a 14-day plan.

Why hospital-style data needs more care than a typical Kaggle file

A retail dataset leaking is a nuisance. A health dataset leaking can expose a diagnosis, a pregnancy, an HIV status or a psychiatric history, and the harm to a person is hard to undo. Three features make health data risky:

• It is sensitive by nature. Diagnoses, medications, test results and procedures reveal intimate details.
• It is easy to re-identify. A rare diagnosis, a small town and an admission date can point to one person even with the name removed.
• It was collected for care, not for your project. Patients agreed to treatment, not necessarily to analysis by a student. Permission to reuse the data has to come from somewhere.

India’s Digital Personal Data Protection Act, 2023 and ethics guidance for health research both bear on this area. Rules and guidance change, so read the current text or ask your institution’s ethics or data-protection contact rather than relying on a summary, including this one.

Which dataset should you use in a two-week internship?

Decide this before touching a single row. The data source determines your obligations.

If someone offers you a raw patient export on a pen drive or by messaging app, stop. Ask who authorised the sharing, under what approval, and where the data may be stored. If nobody can answer, decline politely and switch to synthetic or public data.

Consent and ethics questions to answer before you clean anything

Cleaning is processing. It counts as using the data, so the ethics questions come first.

1. Who owns or controls this data? Name the person or department.
2. What was the original consent? Did patients consent to research use, or was consent waived by an ethics committee? Get the answer in writing.
3. Is there ethics approval? Ask for the approval reference, or confirmation that the data is exempt because it is already anonymous.
4. What is the permitted purpose? Learning exercises and published research are different purposes. Check that yours is covered.
5. Where may the data live? Personal laptop, shared cloud drive and college server are not equivalent.
6. Who else may see it? Mentors, teammates and reviewers each need to be allowed.
7. When must it be deleted? Fix a date and a method.
8. Who do you tell if something leaks? Know the contact before you need it.

The de-identification checklist: from raw columns to a shareable file

Work through these in order and tick each one. Keep the original file untouched and in restricted storage, and clean a copy.

Step A: Inventory the columns

• List every column and label it: direct identifier, indirect identifier (quasi-identifier), sensitive attribute, or neutral.
• Check free-text columns separately. Doctor notes often contain names, phone numbers and places.
• Check file names, sheet names, metadata and comments for identifying text.

Step B: Remove direct identifiers

• Names, including relatives’ names.
• Phone numbers, email addresses, full postal addresses.
• Government ID numbers (such as Aadhaar), hospital registration numbers, medical record numbers, insurance policy numbers.
• Photographs, signatures, biometric data, device identifiers.

Step C: Reduce indirect identifiers

• Replace exact date of birth with age, or an age band such as 30 to 39.
• Shift or generalise admission and discharge dates (for example, keep month and year, or use “days since admission”).
• Replace PIN codes or village names with a broader region.
• Group rare categories (a rare diagnosis, an unusual occupation) into “Other”.
• Cap extreme values such as age above 90 into a single “90+” group.

Step D: Pseudonymise carefully

• If you need to link records of the same patient, replace the ID with a random token, not a transformation of the original ID.
• Do not hash a phone number or record number and treat it as safe. Short, structured numbers can be guessed and re-hashed.
• Store any key that maps tokens back to people separately from the dataset, and do not take it into the analysis environment.

Step E: Test for re-identification

• Pick the quasi-identifiers (for example age band, sex, region, diagnosis).
• Count how many rows share each combination. Any group of one or two rows is a risk.
• Merge, generalise or suppress risky groups until every combination has a minimum group size agreed with your mentor (a common practice is a threshold such as five, but confirm what applies to you).

Step F: Document and store

• Record every transformation in a change log.
• Save the clean file under a clear name, such as patients_deid_v1.csv.
• Protect files with access controls, and do not email data or paste it into public online tools, including AI chatbots.

A small illustrative example: re-identification hiding in a “clean” table

The following is invented for illustration. No code was run and no real patients are involved.

Suppose a student removes names and phone numbers from a table and keeps four columns: age, sex, town and diagnosis. Counting rows per combination might give this:

The first row is a group of one. Anyone who knows a 34-year-old woman in that town with that condition can confirm it is her. The fix is to generalise:

• Change age 34 to the band 30 to 39.
• Change “Small town A” to a district or broader region.
• Group “Rare condition X” into “Other chronic conditions” for the shared version.

Recount after each change. You are done only when no small groups remain, and you note the trade-off: broader categories lose some analytical detail, which is acceptable and should be stated in your report.

Cleaning tasks that are safe to do after de-identification

Once identifiers are handled, normal data cleaning can proceed, though a few health-specific points apply:

• Standardise units. Glucose in mg/dL and mmol/L, weight in kg and lb, and temperature in Celsius and Fahrenheit often mix in one column.
• Handle impossible values. A heart rate of 0 or an age of 240 is an entry error. Flag it, do not silently delete it.
• Treat missingness as information. A missing lab value may mean the test was not ordered, which is clinically meaningful. Do not fill every blank with the mean.
• Normalise categories. “DM2”, “T2DM” and “Type II diabetes” may need one label, ideally agreed with a domain expert.
• Keep a log. Each rule, such as “ages above 120 set to missing”, goes in the change log.
• Avoid clinical conclusions. Describe patterns. Do not give medical advice or diagnoses from your analysis.

Sample data-handling note you can read, adapt and sign off

Adapt this to your project and have your mentor or data owner countersign. It is a learning template, not legal advice, and it does not replace any approval your institution requires.

DATA-HANDLING NOTE (learning project)

Project title: ______________________
Analyst name: _______________________
Mentor / data owner: ________________
Date: _______________________________

1. Data source
   [ ] Synthetic   [ ] Public, de-identified   [ ] Approved project data
   Source name and licence/approval reference: ______________

2. Permission
   Consent basis or ethics approval/exemption reference: ______________
   Permitted purpose: ______________

3. Direct identifiers removed
   [ ] Names  [ ] Contacts  [ ] Addresses  [ ] ID numbers
   [ ] Free-text fields checked  [ ] File metadata checked

4. Indirect identifiers reduced
   Methods used (age bands, date generalisation, region grouping,
   rare-category merge): ______________
   Minimum group size achieved: ______

5. Storage and sharing
   Location: ______________  Access limited to: ______________
   No data sent to public online tools: [ ] Confirmed

6. Retention
   Delete original and working copies by: ______________

7. Incident contact: ______________

I confirm I have followed the steps above and will not attempt to
re-identify any individual.

Analyst signature: ____________   Mentor signature: ____________

Dr. D.Y. Patil Vidyapeeth students: where health-data ethics fits and what to confirm

Dr. D.Y. Patil Vidyapeeth in Pimpri, Pune, was established in 2003 as a Deemed-to-be-University, and it runs its constituent institutes under one umbrella. According to the university’s placements listing, these include Medical, Dental, Biotechnology and Bioinformatics, Nursing, Physiotherapy, Allied Health Sciences, Science and Technology, and a Centre for Online Learning, among others. The careers page describes a NAAC 3rd-cycle accreditation at A++ grade.

What this means in practice: a data analytics student at DPU may sit close to clinical and life-science colleagues, so the chance to work on health-style data is real. That closeness is also why you should be strict about permissions. The published pages reviewed here do not set out internship, project-work or data-handling rules, and it is not confirmed whether such rules are set university-wide or by each institute. The university site also lists a Manual of Examinations page and a syllabus listing page, but check the current versions yourself, because the listed material may or may not address projects or internships. Likewise, the academic calendars page links yearly PDFs through 2026-27, so check your own semester dates before you plan a 14-day block.

Verification steps:

1. Ask your department coordinator in writing whether external online internships are permitted and whether they carry any academic recognition at your institute.
2. Ask whether your institute has an ethics committee or data-protection contact, and what it requires before students handle health-related data.
3. If a faculty member offers project data, request the approval or exemption reference and the permitted-use statement before you open the file.
4. Check your current syllabus and the Manual of Examinations for project or internship evaluation rules that could affect how you document this work.
5. Compare the 2026-27 calendar for your institute against the 14 days you plan to use.

A 14-day plan for a privacy-first health data cleaning project

The free two-week online internship at CareerSwami offers a free offer letter and a free completion certificate after successful completion, across 26 tracks. If you pick a data analytics track, this plan shows how a health-data mini project could fit into 14 days. Confirm the exact deliverables with your track instructions.

Warning signs that a health dataset is not ready for a student project

• It arrives with real names, phone numbers or hospital numbers and nobody can explain the approval.
• You are asked to use a personal email or messaging app to receive or return files.
• The data owner cannot tell you what patients consented to.
• You are told “it is only for practice, so it does not matter”.
• Free-text clinical notes are included and no one has checked them.
• The dataset is tiny and covers a rare condition in a small area, which makes re-identification likely even after cleaning.

How to present this work to a recruiter or faculty reviewer

Privacy practice can differentiate a portfolio. Show the change log, the before-and-after risk table and a redacted version of the data-handling note. Describe the method rather than any real patient detail. Do not claim clinical insight, and do not claim the project is approved by any institution unless you hold written confirmation.

Frequently Asked Questions

Is removing the patient name enough to anonymise a dataset?

No. Combinations of age, sex, location, dates and diagnosis can still single someone out. You also need to generalise indirect identifiers and test group sizes.

Can I use real hospital data in a two-week online internship?

It is rarely practical. Real data needs approval, a consent basis or waiver, and secure storage. Synthetic or public de-identified data lets you build the same skills without those risks.

Is hashing patient IDs a safe way to pseudonymise?

Not by itself. If the original numbers follow a predictable format, they can be guessed and hashed again to match. Use random tokens and keep any mapping key separate and protected.

Can I paste a few sample rows into an AI chatbot to help me clean them?

Not if they are real or possibly real patient records. Public tools may store inputs. Use synthetic rows or describe the column structure without sharing values.

What is a quasi-identifier?

It is a field that is not unique alone but can identify someone when combined with others, such as age, PIN code, sex, admission date or a rare diagnosis.

Does this checklist replace my college’s ethics approval process?

No. It is an independent learning guide. If your project involves data from your institution, follow its approval process and confirm requirements in writing with the relevant faculty or committee.

Does the CareerSwami internship include a stipend or academic credit?

No such claim is made. The confirmed benefits are a free offer letter and a free completion certificate after successful completion. Ask your own institute whether it recognises any external internship for credit.

What should I do if I realise I have received identifiable data by mistake?

Stop working with it, do not copy or forward it, and tell your mentor or the data owner straight away. Follow their instructions on secure deletion and keep a short written record of what happened.

Related Data Analytics guides

• Shri Khushal Das University: A Four-Week Data Analytics Internship Roadmap
• Murshidabad University Students: Running a Data Analytics Internship on a Phone and Weak Network
• JUIT Students: Turning a Python Pandas Notebook into a Reviewable Internship Deliverable
• All Data Analytics internship guides