pISSN 2234-3806
eISSN 2234-3814

Table. 1.

Table 1
Current and desired states of laboratory data and AI and ML modeling

Source of bias	Current state	Desired state	Ongoing initiatives	Gaps/opportunities
Application of cutoff values to non-harmonized test methodologies	Methods used to derive cutoff values in clinical practice guidelines are seldom reported	Methods used to derive cutoff values are disclosed and transparent to allow applicability determination		Journals should require reporting of instrumentation and methods used in clinical trials and clinical practice guidelines
Lack of harmonization of laboratory tests	Only a subset of laboratory tests has been harmonized by manufacturers to produce comparable results across platforms	The majority of in-vitro diagnostic tests have undergone harmonization by the manufacturer	International Consortium for Harmonization of Clinical Laboratory Results	Increase reference material available for harmonization efforts
Instrument and method not reported with laboratory results	Laboratory results do not include the method or instrument used to derive the results, limiting result comparability evaluation of non-harmonized tests	Laboratory results are encoded with instrument and reagent kit identifier	SHIELD	Dissemination and uptake of standard ontology recommendations from SHIELD; EHR and laboratory system functionality to support standard ontology recommendations
Lack of standardization in the digital representation of laboratory results	Variability in how tests are named and results are reported	Accurate representation of laboratory test results using standard ontologies	SHIELD	Dissemination and uptake of standard ontology recommendations from SHIELD; EHR and laboratory system functionality to support standard ontology recommendations
Degradation of Artificial Intelligence Models when applied to different datasets and changing data representation	AI and ML models may lack generalizability when applied to a setting with different data representations and different result values due to the use of non-harmonized tests	Local evaluations are conducted to ensure that the model performs as expected upon retraining or fine-tuning the model with local data as needed, along with performance monitoring over time	CHAI	Dissemination and uptake of recommendations
Insufficient representation of women and minority populations in datasets and clinical trial results	Clinical trials do not routinely report results by sex, race, or ethnicity	Datasets include sex, race, or ethnicity in outcome reports and as a covariate in statistical analysis	NIH	Increased compliance with National Institutes of Health policies
			Revitalization Act of 1993
Bias in AI and ML models because of erroneous race data collected and conclusions inferred in the healthcare literature	Predictive models using race are vulnerable to bias that may perpetuate health disparities because of inaccurate race representation and sufficient data on minorities in datasets	Developers are mindful of potentially erroneous race data collection and inaccurately inferred conclusions and employ the concept of “counterfactual fairness” to ensure models do not unfairly disadvantage minority populations	The Alan Turing Institute Counterfactual Fairness Project	Dissemination and uptake of recommendations

Abbreviations: AI: artificial intelligence; ML: machine learning; SHIELD: Systemic Harmonization and Interoperability Enhancement for Laboratory Data; CHAI: Coalition for Health AI; NIH: National Institutes of Health; EHR, electronic health record.

Ann Lab Med 2025;45:12~21 https://doi.org/10.3343/alm.2024.0323