Evidence of Unreliable Data and Poor Data Provenance in Clinical Prediction Model Research and Clinical Practice

This paper warns that widely used clinical prediction models based on two popular Kaggle datasets lack verifiable data provenance and appear to be fabricated, leading to potentially unreliable research and clinical applications, and calls for mandatory data provenance reporting to safeguard patient care.

The Gist

Imagine you are a chef trying to create a new, life-saving recipe for a soup that cures a specific illness. To do this, you need a cookbook with high-quality, verified ingredients.

This paper is essentially a food safety inspection that discovered a terrifying problem: many chefs are using a "cookbook" found on a popular online cooking forum (Kaggle) that turns out to be filled with fake ingredients.

Here is the breakdown of what the researchers found, using simple analogies:

1. The "Fake Ingredient" Problem

The researchers looked at two very popular datasets (collections of information) on Kaggle, a website where people share data to practice their computer skills. One dataset was about strokes, and the other was about diabetes.

• The Analogy: Imagine these datasets are like a bag of flour sold at a market. The label says "100% Real Wheat Flour." But when you look closely, the flour is actually just chalk dust and sand mixed together.
• The Evidence: The researchers found that the data looked "too perfect." Real patient data is messy; people forget to fill out forms, numbers get lost, and there are gaps. These datasets had almost no missing data, and the numbers followed strange, robotic patterns (like a computer program guessing the numbers rather than a doctor recording them).
• The Source: The people who uploaded these files admitted they couldn't tell anyone where the data came from. One even said, "Don't use this for research!" Yet, people did anyway.

2. The "Fast-Food" Research Factory

The paper describes a trend called "fast-churn research."

• The Analogy: Imagine a factory that wants to make as many burgers as possible to get rich, so they stop checking if the meat is safe. They just slap a bun on a rock, call it a burger, and sell it.
• The Reality: Researchers are downloading these fake datasets, running computer programs on them, and publishing papers very quickly to get famous or get grants. They aren't trying to help patients; they are just trying to publish papers.

3. The Dangerous Consequences

This is where the story gets scary. Because these "fake ingredient" papers were published in scientific journals, other people started trusting them.

• The Domino Effect:
  • 124 Papers: 124 different studies were written using this fake data.
  • 86 Reviews: These bad papers were cited in 86 other review articles, spreading the misinformation like a virus.
  • Real-World Use: Even worse, three of these models are actually being used in hospitals or medical devices to make decisions about real patients.
• The Risk: If a doctor uses a model built on fake data, they might tell a patient they are healthy when they are sick, or vice versa. It's like a GPS built on a map of a city that doesn't exist; it will confidently drive you off a cliff.

4. The "No Receipt" Policy

The paper points out that the "supermarkets" (Kaggle) and the "restaurants" (Journals) have no rules about checking receipts.

• The Analogy: If you buy a car, you expect a VIN number and a history report. But in medical research, you can buy a dataset with no history, no owner, and no proof of where it came from.
• The Failure: The researchers checked the "receipts" (data provenance) for these two datasets using a strict checklist called TRIPOD+AI. Both datasets failed every single item. They had no information on who collected the data, when, where, or why.

5. The Solution: A New "Health Code"

The authors aren't just pointing fingers; they are proposing a new set of rules to fix the kitchen:

• For the Supermarkets (Kaggle): They must force uploaders to fill out a "Provenance Sheet" (a detailed receipt) that says exactly where the data came from. If you can't prove it's real flour, you can't sell it.
• For the Restaurants (Journals): They shouldn't accept a paper unless the chef shows the raw ingredients first. If the ingredients look fake, reject the paper immediately.
• For the Chefs (Researchers): Stop trying to be fast. Check your ingredients. If a dataset looks too good to be true, it probably is.

The Bottom Line

This paper is a wake-up call. It says that we cannot trust our medical predictions if the foundation they are built on is made of sand.

If we want to save lives with AI and data, we need to stop using "fake flour" and start demanding verified, real ingredients. Otherwise, we risk building a hospital on a foundation that will collapse the moment a real patient walks through the door.

Technical Summary

1. Problem Statement

The paper addresses a critical integrity issue in medical research: the proliferation of clinical prediction models (CPMs) developed using large, publicly available datasets that lack data provenance (information regarding the origin, collection, and processing of data).

• The Context: There is a trend of "fast-churn" research where authors rapidly publish papers using public datasets to generate citations, often without verifying data quality or authenticity.
• The Specific Issue: The authors identified two highly popular datasets on Kaggle (a platform for data science competitions) regarding stroke and diabetes prediction. These datasets are widely used in peer-reviewed literature but have no verifiable source, no ethical clearance documentation, and exhibit statistical anomalies suggesting they are simulated or fabricated.
• The Risk: Models built on such data are being cited in reviews, used in medical device patents, and potentially deployed in clinical practice, posing a direct risk of patient harm through inaccurate predictions.

2. Methodology

The study employed a mixed-methods approach combining data forensics, bibliometric analysis, and guideline adherence assessment.

• Dataset Selection: Two Kaggle datasets were selected:
  1. Stroke Dataset: 5,110 patients, 12 variables.
  2. Diabetes Dataset: 100,000 patients, 9 variables.
• Provenance Assessment: The datasets were evaluated against nine specific items from the TRIPOD+AI statement (an updated guideline for reporting prediction models), focusing on data origin, collection methods, and preprocessing.
• Exploratory Data Analysis (Forensics): The authors performed statistical checks to detect signs of fabrication:
  • Analysis of correlations between variables (looking for expected vs. unexpected associations).
  • Inspection of variable distributions (checking for uniformity or artificial patterns).
  • Identification of duplicate rows and missing data patterns (real-world health data usually has significant missingness; these datasets had almost none).
  • Correlation between patient IDs and clinical variables (a common sign of synthetic generation).
• Literature Review:
  • A search of Google Scholar identified 124 peer-reviewed articles using these datasets for CPM development or validation.
  • Articles were screened for ethical statements, claims of clinical utility, and adherence to TRIPOD+AI reporting standards.
  • Citation analysis was conducted using OpenAlex, Altmetric, and Overton to track influence on reviews, policy, and patents.

3. Key Contributions

• Identification of "Poisoned" Data Sources: The paper provides concrete evidence that two widely used Kaggle datasets are likely synthetic/fabricated and are currently fueling a significant volume of low-quality research.
• Quantification of Impact: The study maps the downstream effects of these unreliable datasets, revealing that 124 studies have been published, cited over 1,500 times, included in 86 review articles, and referenced in a medical device patent.
• Forensic Methodology: It demonstrates a reproducible framework for detecting synthetic data in health research (e.g., analyzing ID-variable correlations and discrete variable distributions).
• Policy Recommendations: The authors propose mandatory Data Provenance Reporting frameworks for journals, publishers, and data repositories (like Kaggle) to prevent future misuse.

4. Key Results

A. Data Provenance and Authenticity

• Provenance Scores: Both datasets scored 0/9 on the TRIPOD+AI data provenance checklist. The uploaders explicitly stated the source could not be disclosed (e.g., "private source," "confidentiality reasons").
• Stroke Dataset Anomalies:
  • Missing Data: Only 0.3% missing values (unrealistic for routine health data).
  • Distribution Shift: A clear, artificial shift in the distribution of average glucose levels occurred exactly at the midpoint of the patient ID variable.
• Diabetes Dataset Anomalies:
  • Discrete Values: Despite 100,000 patients, blood glucose and HbA1c levels only took 18 discrete values, indicating resampling or fabrication.
  • Duplicates: 7% of patient observations were exact duplicates.
  • Correlations: Abnormal lack of association between BMI and blood glucose/HbA1c, contradicting established medical physiology.

B. Literature Analysis (124 Articles)

• Geographic Spread: Authors from 32 countries used these datasets, with India (31 articles), Indonesia (15), and China (11) being the top contributors.
• Ethical Reporting:
  • 90% of articles had no ethical statement.
  • Only 2% (n=3) claimed ethical approval; 7% claimed ethics were not required.
• Clinical Claims:
  • 68% of stroke articles and 80% of diabetes articles made practical recommendations for clinical use.
  • Three models were explicitly claimed to be used in clinical practice or tested on patients.
  • Two models were developed into publicly accessible web/mobile tools.
  • One article was cited in a medical device patent (Caltech/USC) for stroke risk determination.
• Citation Impact: The articles received 1,529 total citations, including 86 review articles, effectively propagating the unreliable data into the broader scientific consensus.

5. Significance and Recommendations

Significance:
The paper highlights a systemic failure in research integrity where the "open data" movement is being exploited by "fast-churn" research. The use of unverified, potentially synthetic data undermines evidence-based medicine. If these models are deployed, they could lead to misdiagnosis, denial of necessary treatment, or unnecessary interventions. The fact that these models are entering patents and clinical workflows suggests the problem is no longer just academic but poses a public health risk.

Recommendations:
The authors propose a multi-stakeholder approach to enforce data integrity:

1. For Journals/Publishers:
   • Mandate Data Provenance Statements (Who, What, When, Where, Why, Funding) for all submissions.
   • Require submission of unprocessed patient-level data at the time of submission for rapid verification.
   • Automatically reject papers using competition datasets unless the authors provide strong justification and proof of data authenticity.
2. For Data Repositories (e.g., Kaggle):
   • Implement mandatory provenance fields for all health/medical datasets (Table 5 in the paper).
   • Flag or remove datasets that cannot verify their source or ethical clearance.
3. For Researchers/Clinicians:
   • Perform "sense checks" on data distributions and correlations before modeling.
   • Prioritize pre-registered studies and those following TRIPOD+AI and PROBAST+AI guidelines.
   • Do not assume published research is error-free; verify data sources independently.

Conclusion:
The study concludes that without mandatory provenance reporting and rigorous data verification, the reliability of clinical prediction models is compromised. Immediate actions, including the potential retraction of studies based on these specific datasets and the removal of associated clinical tools, are recommended to safeguard patient safety.