PaperlayerView exact revision on arXiv

Source-linked AI summary

The METRIC-framework for assessing data quality for trustworthy AI in medicine: a systematic review

Daniel Schwabe, Katinka Becker, Martin Seyferth, Andreas Klaß, Tobias Schäffter

arXiv:2402.13635v1cs.LGcs.AI

TL;DR

ML research often focuses on a few task-specific data-quality dimensions, a potential shortcoming for safety-critical applications such as medical diagnosis. This paper reviews existing data-quality research for AI in medicine and proposes METRIC, a framework of 15 awareness dimensions for medical training data.

  • Problem

    ML research often focuses on few task-specific data-quality dimensions, a potential shortcoming for safety-critical applications such as medical diagnosis.

    Source: ML data quality · p. 6

  • Method

    The authors conduct a systematic review to transfer existing data-quality research to AI in medicine and examine dataset characteristics for that context.

    Source: Introduction · p. 3

  • Results

    METRIC organizes 15 data-quality dimensions for medical training data as task-specific awareness dimensions for developers.

    Source: Discussion · pp. 11, 13

  • Takeaways & Limitations

    Systematic assessment of medical datasets could serve as a starting point for regulation and has potential to accelerate approval of ML products.

    Source: Discussion · p. 13

    The results are limited to ML approaches, despite the paper's use of broader AI terminology.

    Source: Limitations · p. 3

Abstract

from arXiv · show

The adoption of machine learning (ML) and, more specifically, deep learning (DL) applications into all major areas of our lives is underway. The development of trustworthy AI is especially important in medicine due to the large implications for patients' lives. While trustworthiness concerns various aspects including ethical, technical and privacy requirements, we focus on the importance of data quality (training/test) in DL. Since data quality dictates the behaviour of ML products, evaluating data quality will play a key part in the regulatory approval of medical AI products. We perform a systematic review following PRISMA guidelines using the databases PubMed and ACM Digital Library. We identify 2362 studies, out of which 62 records fulfil our eligibility criteria. From this literature, we synthesise the existing knowledge on data quality frameworks and combine it with the perspective of ML applications in medicine. As a result, we propose the METRIC-framework, a specialised data quality framework for medical training data comprising 15 awareness dimensions, along which developers of medical ML applications should investigate a dataset. This knowledge helps to reduce biases as a major source of unfairness, increase robustness, facilitate interpretability and thus lays the foundation for trustworthy AI in medicine. Incorporating such systematic assessment of medical datasets into regulatory approval processes has the potential to accelerate the approval of ML products and builds the basis for new standards.

Introduction

The paper argues that training-data quality is foundational to trustworthy medical ML and introduces METRIC, a 15-dimension framework for assessing medical training data. Its scope is the content of a fixed dataset for a specific application, and its findings are limited to ML approaches.

Introduction

  • Medical ML is expanding, but its consequences for decision-making and patient lives make reliable health data and more stringent regulatory requirements important for supporting evidence and approvals.Despite reported human-level performance, medical ML products are mainly used as diagnosis assistance, with final decisions left to medical professionals.Source: Introduction · p. 2
  • Training-data quality fundamentally shapes AI behaviour: poor data can yield poor systems, and biased data can cause neural networks to learn and amplify bias, making dataset assessment critical to trustworthy AI.The paper identifies data quality as decisive for developers, regulators, and notified bodies.Source: Introduction · pp. 2–3
  • The review combines prior data-quality research with modern ML perspectives to propose METRIC, a framework of 15 awareness dimensions for medical training data.It is intended to help developers understand how dataset composition affects ML application behaviour and trustworthiness.Source: Introduction · p. 3

Limitations

  • METRIC assesses dataset content for a fixed dataset and specific application, excluding data-governance and data-management frameworks, survey-data case studies, assessment in a vacuum, and strategies for coping with bad data.The focus is content used by ML to learn patterns, rather than surrounding technical infrastructure.Source: Limitations · p. 3
  • Although the paper uses “AI” and “trustworthy AI” as common terms, its results are limited to machine-learning approaches, particularly deep learning.The authors note that healthcare discussions using “AI” generally concern ML and DL, although AI encompasses a wider range of technologies.Source: Limitations · p. 3

Results

The review included 62 studies spanning general, big, and ML data quality, and found that ML research largely examines isolated data-quality dimensions and predictive robustness rather than comprehensive frameworks. Its synthesis presents METRIC as a medical training-data framework with 15 dimensions grouped into five clusters.

Results

  • The review included 62 of 2,362 identified studies; its corpus comprised 28 general-data, 7 big-data, and 27 ML-data entries.The studies span roughly 30 years of data-quality research.Source: Figure 1 · p. 4, Results · p. 4, Figure 2 · p. 5

General data quality

  • Medical data-quality frameworks evolved from EHR-focused approaches toward data-type-specific frameworks, but population representativeness and consistent terminology remained challenges.EHR frameworks addressed accuracy, completeness, and timeliness, while later work sought harmonised definitions and assessment.Source: General data quality · p. 5

Big data quality

  • Big-data frameworks addressed schema and infrastructure efficiency alongside data quality, while the 4 V’s characterize big data rather than define quality dimensions.The review notes that computational capability can constrain the benefits of using more data.Source: Big data quality · pp. 5–6

ML data quality

  • ML studies mostly tested individual dimensions: more data improved performance with saturation, while completeness and errors were studied as missing data, contaminated inputs, or noisy targets.Studies commonly manipulated training data, such as distorting image contrast or replacing correct labels with false ones.Source: ML data quality · p. 6
  • Most ML studies assessed robustness through predictive performance; generalisability and predictive uncertainty received less attention, while explainability and fairness were underrepresented.The review identifies this imbalance as a potential shortcoming for safety-critical applications such as medical diagnosis prediction.Source: ML data quality · p. 6, Figure 3 · p. 7
  • METRIC presents a specialised framework for medical training data with 15 data-quality dimensions organized into five clusters.The framework is depicted as a wheel of data quality.Source: Figure 3 · p. 7

METRIC-framework for medical training data

The METRIC-framework organizes medical training-data quality into purpose-specific awareness dimensions, separating dataset-content assessment from data-management prerequisites. It emphasizes assessing acquisition errors, timeliness, representativeness, informativeness, and consistency for a known medical-AI use.

METRIC-framework for medical training data

  • METRIC is a purpose-specific framework of awareness dimensions, not a measurement guideline, for assessing medical training data against a known medical-AI use.It organizes 38 dimensions and subdimensions into clusters, dimensions, and subdimensions; a separate data-management cluster is a prerequisite for content-quality assessment.Source: METRIC-framework for medical training… · pp. 7–8

Measurement process

  • Measurement-process assessment covers device and human errors, missing values, source credibility, and data poisoning because errors can propagate to false diagnoses or treatment.Accuracy may be impossible to evaluate when medical ground truth is unavailable; training noise should then be compared with expected deployment noise, and adding noise may sometimes improve performance. Human label noise and observer variability also warrant attention.Source: Measurement process · pp. 9–10

Timeliness

  • Timeliness assessment checks whether data age and updates suit the task, since evolving medical knowledge and coding systems can make labels or standards outdated.Data age may also indicate the measurement technology used, informing assessment of precision and accuracy.Source: Timeliness · p. 10

Representativeness

  • Representativeness assessment examines demographic and source variety, data depth and subpopulation coverage, and class balance to judge applicability to the target population.Rare classes may need deliberate overrepresentation rather than matching real-world prevalence so ML algorithms can learn their patterns.Source: Representativeness · pp. 10–11

Informativeness

  • Informativeness assessment considers understandability, redundancy, informative missingness, and feature importance to determine whether data conveys clear, useful information for the task.Missingness patterns can correlate with diagnoses, while valuable features may matter as much as dataset size given computational constraints.Source: Informativeness · p. 11

Consistency

  • Consistency assessment checks rule-based format and compliance, logical contradictions, and statistical consistency across subsets and time points.Distribution drift may be omitted for datasets that do not change continuously, but lack of model surveillance can cause it to be overlooked.Source: Consistency · p. 11

Discussion

The METRIC-framework offers 15 awareness dimensions to guide developers in understanding medical training data and its effects on AI behavior. The discussion outlines how to assess and prioritize those dimensions and how systematic dataset assessment could support trustworthy AI and regulation.

Discussion

  • The 15 METRIC awareness dimensions guide developers in understanding dataset characteristics and how they shape AI behavior, helping improve data acquisition and selection to reduce bias, increase robustness, and facilitate interpretability.These improvements have the potential to strengthen AI trustworthiness.Source: Discussion · pp. 11–12
  • Overall data-quality assessment requires choosing a measure for each dimension, obtaining a result, and judging its appropriateness for the task; combining outcomes may support an overall measure.The framework currently divides data quality into dimensions, but each dimension and subdimension still needs suitable quantitative or qualitative measures and task-specific evaluation.Source: Discussion · p. 12
  • The dimensions are unranked, although some may warrant early assessment or dataset exclusion, while qualitative measures, complex statistics, use-case dependence, and required expertise can make assessment difficult.These considerations matter for developing a complete data-quality assessment and examination process.Source: Discussion · p. 12
  • About half the dimensions mostly require quantitative measures and a fifth mostly qualitative ones; representativeness, timeliness, device error, and feature importance are use-case dependent.Quantitative measures can support objectivity and automation, while qualitative or mixed assessment may require difficult-to-obtain, costly medical domain knowledge. Figure 5 categorizes dimensions by measurement type and use-case dependence.Source: Discussion · pp. 12–13, Figure 5 · p. 12
  • The authors recommend prioritizing six dimensions: the representativeness cluster, feature importance, distribution consistency, and human-induced error; all but feature importance are mostly measured quantitatively.These dimensions may have proportionally larger effects on AI applications, though the effects may depend on the ML problem and remain to be quantified; the mostly quantitative dimensions are candidates for software tools.Source: Discussion · p. 13
  • Systematic medical-dataset assessment using METRIC and its additional considerations could inform regulation and certification guidelines and potentially accelerate approval of medical ML products.The discussion anticipates that future guidelines may evaluate training and testing dataset quality alongside AI algorithms.Source: Discussion · p. 13

Methods

The authors conducted a PRISMA-guided review of data-quality frameworks and training-data quality in medical ML, using database searches, screening, and snowballing to select records. They extracted and grouped data-quality concepts and mapped literature-reported terms to METRIC dimensions, definitions, and references.

Literature review

  • The review followed PRISMA guidelines, screening titles and abstracts, snowballing references, and assessing remaining full texts against eligibility criteria.Two authors independently screened titles and abstracts, resolving disagreements through discussion or consultation with a third author.Source: Literature review · p. 14, Literature review process · p. 15

Search strategy

  • The search combined data-quality terms with framework and dimension terms, and with machine-learning, artificial-intelligence, and deep-learning terms; PubMed and ACM Digital Library results were concatenated and deduplicated, yielding 1,735 records.Source: Search strategy · p. 14

Search query

  • The search used exact keyword matching in database title and abstract fields.Keywords were enclosed in quotation marks, rather than searched fuzzily.Source: Search query · p. 14

Eligibility criteria

  • Papers qualified if they offered broad-scale general or medical data-quality frameworks or examined a training-data dimension’s effect on DL behavior; non-English and inaccessible records were excluded.The review imposed no publication-date or publisher-source limits.Source: Eligibility criteria · p. 15, Table S2 · p. 15

Literature review process

  • Two authors independently screened database records against the eligibility criteria, resolved disagreements by discussion, and consulted a third author if needed.This screening reduced the database-search results to 104 records.Source: Literature review process · p. 15
  • 62 entries remained after full-text screening of 182 papers, following database and snowballing searches.The database search produced 104 records after title-and-abstract screening; snowballing added 78 eligible papers from 585 references.Source: Literature review process · p. 15

Data extraction strategy

  • Two authors read each of the 62 selected records and extracted relevant data-quality terms, excluding source-specific, infrastructure-focused, or uninterpretable terms.Terms were recorded in an Excel sheet as the starting template for METRIC; subdimensions generally required mention by at least three corpus references, with variety of data sources retained as an exception.Source: Data extraction strategy · pp. 15–16
  • Authors grouped subdimensions hierarchically into dimensions and clusters and reached consensus on definitions, drawing on glossary, Wang et al., or the reviewed literature.The definitions were framed for medical training data.Source: Data extraction strategy · p. 16
  • Extracted dimensions include coverage and representativeness, completeness, currency or timeliness, and data granularity.Source: Data extraction strategy · pp. 17–18
  • The extraction table organizes quality information by measurement process, timeliness, representativeness, informativeness, and consistency.Source: Data extraction strategy · pp. 18–19
  • The coded quality concerns include disagreement among repeated observations, noisy labels, carelessness, and data poisoning.Source: Data extraction strategy · p. 18
  • Other extracted dimensions include uniqueness, syntactic and semantic consistency, plausibility, and distribution drift.Source: Data extraction strategy · p. 19
  • The extraction table records datatype, year, author, and measurement, spanning dimensions such as timeliness, representativeness, informativeness, consistency, completeness, and accuracy.Source: Data extraction strategy · pp. 19–20
  • The extracted ML-related dimensions include dataset size, coverage, target-class balance, noisy labels, and noisy attributes.Source: Data extraction strategy · pp. 19–20
  • Other cataloged concerns include image noise and distortions, distribution drift, missing values, duplicates, conflicting data, and outliers.Source: Data extraction strategy · p. 20
  • The corpus table records publication metadata and maps each paper’s reported data-quality terms to METRIC dimensions and subdimensions across five clusters.It also categorizes datatype, publication year, authors, and field.Source: Table S3 · p. 21

Definitions of data quality dimensions

  • Supplementary tables define the hierarchy and literature references for Measurement Process, Timeliness, Representativeness, Informativeness, and Consistency, plus prerequisite Data Management.Each cluster table provides term definitions and connected corpus references.Source: Table S4 · p. 22, Table S5 · p. 23, Table S6 · p. 23

Competing interests

The authors declare no competing interests. They acknowledge EU TEF-Health project funding and provide the extracted literature-corpus data and terms in a supplementary Excel file.

Competing interests

Funding

  • The work was funded by the EU TEF-Health project, part of the EU Digital Europe Programme.The funding identifier is DIGITAL-2022-CLOUD-AI-02-TEF-HEALTH.Source: Funding · p. 39

Additional information

  • Extracted literature-corpus data and terms underlying the METRIC-framework are available in a supplementary Excel file.Source: Additional information · p. 39

Related papers

Loading 2402.13635v1…