Availability and Quality of Anthropometric Data in Swiss Childrens Hospitals: The SwissPedGrowth Project

The SwissPedGrowth project demonstrates the feasibility of extracting high-quality anthropometric data from heterogeneous Swiss children's hospital electronic health records for growth research, despite challenges regarding data completeness and the need for weighting to ensure population representativeness.

The Gist

Imagine you are trying to build a giant, perfect map of how children in Switzerland grow up. To do this, you need two things: a lot of data points (heights and weights) and a way to make sure those points are accurate.

This paper, called SwissPedGrowth, is like a report card on a massive, ambitious attempt to build that map using the digital medical records (Electronic Health Records or EHRs) from seven major children's hospitals across Switzerland.

Here is the story of their journey, explained simply:

1. The Big Idea: The Digital Treasure Hunt

The researchers wanted to see if they could dig through the digital filing cabinets of seven different hospitals (in cities like Zurich, Geneva, and Bern) to pull out height and weight measurements for nearly half a million kids.

Think of these hospitals as seven different libraries. Each library has its own unique way of organizing books (data). Some use card catalogs, some use digital databases, and some have notes scribbled on napkins. The team wanted to see if they could create a "universal translator" to read all these different systems at once and pull out the growth data.

2. The Challenge: The "Messy Attic" Problem

The first hurdle was finding the data.

• The Reality: They found that while doctors did measure the kids, the data wasn't always easy to find. It was like looking for a specific sock in a messy attic.
• The Results: They found weight measurements in about 43% of visits, but height measurements in only 20%.
• Why? Often, doctors wrote the numbers in unstructured text (like a free-form note) or on scanned paper documents that computers couldn't "read." It's like trying to find a specific sentence in a book where the author wrote it in invisible ink or on a sticky note stuck to the back cover.

3. The Quality Control: The "Detective Work"

Once they found the numbers, they had to check if they were real.

• The Problem: Computers are literal. If a doctor accidentally typed "150" for a baby's weight (thinking of grams) instead of "1.5" (kilograms), the computer thinks the baby weighs as much as a grand piano! Or, a doctor might copy-paste a measurement from last year's visit without checking if it's still true.
• The Solution: The team built two "detective bots."
  1. Bot A (Their own creation): Looked for numbers that made no biological sense (e.g., a 5-year-old who is 3 meters tall).
  2. Bot B (An existing tool): Checked if the numbers changed logically over time.
• The Outcome: They caught some funny errors (like swapped height/weight numbers) and realized that about 30% of the height data was just "copy-pasted" from previous visits. They cleaned up the mess, throwing out the "bad apples" so the remaining data was trustworthy.

4. The Big Picture: Is This Group of Kids "Normal"?

A major worry was: Are these kids in the hospital records just a weird, sick group, or do they represent all Swiss children?

• The Comparison: They compared their group of 477,000 kids to the entire population of Swiss children under 20.
• The Findings: At first glance, the hospital kids looked slightly different (a bit younger, slightly more boys, and from slightly wealthier neighborhoods).
• The Fix: They used a statistical "magic trick" called weighting. Imagine you have a bag of marbles that is too heavy on one side. You add a few light marbles to the other side to balance it out. After this balancing act, the hospital group looked almost identical to the general Swiss population. This means the data can be used to understand all Swiss children, not just the ones in the hospital.

5. The Verdict: It's Possible, But Hard Work

The paper concludes that yes, we can use hospital computer records to study how children grow, but it is hard work.

• The Good News: The data exists, it's high quality once cleaned, and it represents the real population.
• The Bad News: It took a huge amount of effort to clean up the data because doctors often write things in ways computers hate (free text, scanned papers).
• The Future: To make this easier next time, doctors need to be trained to type data into specific boxes (like filling out a form) rather than writing free notes, and hospital computer systems need to get smarter at organizing this information.

The Takeaway

Think of this project as building a bridge across a river. The river is the gap between "what doctors write down" and "what researchers can use." The SwissPedGrowth team successfully built the bridge, but they had to spend a lot of time clearing rocks and debris (cleaning the data) to make sure it was safe to cross. Now that the bridge is open, researchers can finally drive their research cars across it to learn more about child health!

Technical Summary

1. Problem Statement

Anthropometric data (height, weight, head circumference) are critical for pediatric care, growth monitoring, and epidemiological research. While these measurements are routinely collected and stored in Electronic Health Records (EHRs), leveraging them for large-scale research faces significant barriers:

• Data Heterogeneity: Swiss children's hospitals utilize diverse EHR systems, making data extraction and harmonization difficult.
• Data Quality & Completeness: There is uncertainty regarding the availability of structured data versus unstructured free text, and the prevalence of errors (e.g., unit errors, duplicates, biologically implausible values).
• Representativeness: It is unclear if clinical cohorts derived from EHRs accurately reflect the general pediatric population in terms of demographics and socioeconomic status.
• Technical Challenges: The transition from raw clinical documentation to research-ready data involves complex mapping to standardized schemas (like SPHN) and handling of administrative vs. clinical visit structures.

2. Methodology

The study, SwissPedGrowth, is a multicenter, retrospective analysis nested within the SwissPedHealth national data stream.

• Data Source & Scope:
  • Hospitals: Seven children's hospitals across Switzerland (Basel, Bern, Geneva, Lausanne, Lucerne, St. Gallen, Zurich).
  • Population: Patients <20 years old with visits between 2017 and 2023.
  • Volume: Initial extraction included 560,995 patients and 2.3 million visits; final analysis included 477,531 patients and 2,171,633 visits after exclusions (missing sex or visit type).
• Data Extraction & Harmonization:
  • Data was extracted from local Clinical Data Warehouses (CDWs) and mapped to the Swiss Personalized Health Network (SPHN) schema using standard terminologies (SNOMED-CT, LOINC).
  • Data was converted to Resource Description Framework (RDF) format and transferred to a secure server (BioMedIT) for processing.
  • Socioeconomic Linkage: The Swiss Neighbourhood Index of Socioeconomic Position (Swiss-SEP) was linked to patient addresses (which were then deleted) to assess socioeconomic status.
  • Visit Classification: Visits were categorized as Outpatient, Emergency, or Inpatient.
• Data Quality Assessment:
  • Algorithms: Three approaches were used to clean data:
    1. A self-developed algorithm based on WHO growth reference z-scores.
    2. The existing growthcleanr algorithm (longitudinal moving average).
    3. A sequential combination of both.
  • Error Detection: The algorithms flagged unit/decimal errors, swapped values (height/weight), duplicates (same-day or carried forward), invalid values (negative/zero), and biologically implausible outliers.
  • Validation: A manual review of 140 visits (70 inpatient, 70 outpatient) where data was not automatically extracted was conducted to identify sources of missing data (e.g., unstructured text, scanned documents).
• Representativeness Analysis:
  • Compared study demographics (age, sex, nationality, Swiss-SEP) against the general Swiss population (<20 years) using census data.
  • Calculated standardized differences (Cohen's h / Glass's $\Delta$) and applied iterative proportional fitting (raking) to weight the study population to match national distributions.

3. Key Contributions

• Feasibility Demonstration: Proved that large-scale, multi-institutional extraction of pediatric anthropometric data is feasible in Switzerland using the SPHN framework, despite heterogeneous EHR systems.
• Algorithmic Innovation: Developed and validated a custom cleaning algorithm that, when combined with growthcleanr, effectively identified and corrected data errors, providing a robust pipeline for future research.
• Granular Quality Metrics: Provided the first comprehensive breakdown of data availability and quality across different visit types (inpatient vs. outpatient vs. emergency) in the Swiss context.
• Representativeness Framework: Demonstrated that while raw EHR cohorts have demographic biases, these can be effectively corrected via statistical weighting to create a population-representative dataset.

4. Key Results

A. Data Availability

• Overall Rates: Height was available in 20% of visits, weight in 43%, head circumference in 5-6%, and calculated BMI in 23%.
• Variation by Visit Type: Availability was significantly higher for inpatient stays compared to outpatient or emergency visits.
  • Reason: Inpatient data is often structured; outpatient data is frequently buried in unstructured free text (e.g., clinical summaries, scanned nursing charts).
• Demographic Bias: Height and weight were more frequently recorded for infants and Swiss nationals compared to older children and non-Swiss nationals.

B. Data Quality

• Error Rates (Combined Algorithms):
  • Invalid Recordings: <1% (zero/negative values).
  • Biologically Implausible Outliers: 4% of heights, 3% of weights.
  • Duplicates/Carried Forward: 29% of heights and 31% of weights were flagged as duplicates or values carried forward from previous visits.
  • Correction: The combined approach corrected <1% of unit/decimal errors.
• Head Circumference: Showed higher vulnerability, with 7% flagged as outliers and 18% as duplicates.
• Manual Review Findings: In 140 reviewed visits where extraction failed:
  • 93% of inpatient visits had weight data, but it was often in scanned PDFs or unstructured fields.
  • Only 35% of outpatient visits had height data, mostly in free-text fields.

C. Representativeness

• Raw Differences: The unweighted cohort was younger, had more boys, fewer Swiss nationals, and higher socioeconomic status (Swiss-SEP) than the general population.
• Effect Sizes: Differences were generally small (Cohen's d < 0.2) to small-to-medium (Age d ~0.4).
• Post-Weighting: After applying raking weights, standardized differences for all variables (age, sex, nationality, Swiss-SEP) dropped to <0.001, indicating the weighted cohort is highly representative of the general Swiss pediatric population.

5. Significance and Implications

• Research Utility: The study confirms that EHRs can serve as a massive, high-quality source for pediatric growth research, provided rigorous cleaning and weighting are applied.
• Systemic Recommendations:
  • Physician Level: Training is needed to encourage structured data entry over free text to improve extractability.
  • Hospital Level: EHR systems must be optimized to capture anthropometrics in structured fields, particularly for high-volume outpatient and emergency settings.
  • Framework Level: The SPHN schema needs adaptation to better link clinical data to specific visits rather than broad administrative cases, which currently complicates data mapping.
• Limitations: The study lacks data from the Italian-speaking part of Switzerland. Additionally, the reliance on structured fields means data for outpatient visits remains incomplete compared to inpatient data.
• Conclusion: Despite the substantial workload required for harmonization, the SwissPedGrowth project establishes a viable pathway for using real-world EHR data to answer both individual-level clinical questions and public health epidemiological questions regarding child growth in Switzerland.