Diagnostic laboratories across India show significant differences in how they classify common medical test results as normal or abnormal, according to an analysis of nearly 795,000 medical examinations.

The findings point to a data-standardisation challenge that could affect insurers, researchers and developers using artificial intelligence (AI) in healthcare.

The analysis covered 794,811 digitised examinations conducted between January 2023 and May 2026, containing more than 20.7 million individual laboratory-result rows from 33,244 diagnostic laboratories.

Across selected clinical decision points for 10 commonly reported tests, the population-weighted disagreement rate in reference-range classification averaged 24.8%.

The analysis was conducted by researcher and technology professional Saif Lakhani as part of an examination of health risks among Indian life-insurance applicants. It found that the same numerical result could receive different classifications depending on the reference interval used by a laboratory.

The differences were particularly marked at certain thresholds. A fasting blood sugar reading of 110 mg/dL had a disagreement rate of 49.7%, while an LDL cholesterol reading of 130 mg/dL showed a 49.9% disagreement rate.

“Two Indian labs looking at the exact same number would disagree on calling it normal or abnormal about one time in four,” Lakhani said, referring to the average across the 10 tests examined.

Data Fragmentation Hinders Healthcare AI

The analysis also identified inconsistencies in how laboratory data is recorded. It found 9,165 distinct representations of test names and 21,223 representations of units.

Tests such as SGPT, ALT and Alanine Aminotransferase, for instance, can refer to the same measurement but appear differently in datasets.

Such variations can complicate the use of medical records in automated systems. Differences in units, test names and reference ranges may lead AI systems to treat equivalent results as separate observations or interpret laboratory reporting practices as differences in patient health.

“The real work is never the algorithm,” Lakhani said. “It’s making the input trustworthy.”

The analysis recommends greater standardisation of laboratory test coding and units, while insurers and researchers combining data from multiple laboratories should account for lab-specific reference ranges.

The findings do not suggest that differences in reference ranges constitute medical errors, as ranges can vary legitimately based on testing methods, equipment and population characteristics. The dataset was also limited to life-insurance applicants and is not representative of the wider Indian population.

Instead, the analysis highlights a broader challenge as healthcare moves towards data-driven systems: digitised medical records are not necessarily comparable across laboratories without further standardisation.