Dataset Information

A framework for evaluating clinical artificial intelligence systems without ground-truth annotations.

ABSTRACT: A clinical artificial intelligence (AI) system is often validated on data withheld during its development. This provides an estimate of its performance upon future deployment on data in the wild; those currently unseen but are expected to be encountered in a clinical setting. However, estimating performance on data in the wild is complicated by distribution shift between data in the wild and withheld data and the absence of ground-truth annotations. Here, we introduce SUDO, a framework for evaluating AI systems on data in the wild. Through experiments on AI systems developed for dermatology images, histopathology patches, and clinical notes, we show that SUDO can identify unreliable predictions, inform the selection of models, and allow for the previously out-of-reach assessment of algorithmic bias for data in the wild without ground-truth annotations. These capabilities can contribute to the deployment of trustworthy and ethical AI systems in medicine.

SUBMITTER: Kiyasseh D

PROVIDER: S-EPMC10902352 | biostudies-literature | 2024 Feb

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

A framework for evaluating clinical artificial intelligence systems without ground-truth annotations.

Kiyasseh Dani D Cohen Aaron A Jiang Chengsheng C Altieri Nicholas N

Nature communications 20240228 1

A clinical artificial intelligence (AI) system is often validated on data withheld during its development. This provides an estimate of its performance upon future deployment on data in the wild; those currently unseen but are expected to be encountered in a clinical setting. However, estimating performance on data in the wild is complicated by distribution shift between data in the wild and withheld data and the absence of ground-truth annotations. Here, we introduce SUDO, a framework for evalu ...[more]

PMID: 38418453

Similar Datasets

Project description:BackgroundClinical decision support systems are designed to utilize medical data, knowledge, and analysis engines and to generate patient-specific assessments or recommendations to health professionals in order to assist decision making. Artificial intelligence-enabled clinical decision support systems aid the decision-making process through an intelligent component. Well-defined evaluation methods are essential to ensure the seamless integration and contribution of these systems to clinical practice.ObjectiveThe purpose of this study was to develop and validate a measurement instrument and test the interrelationships of evaluation variables for an artificial intelligence-enabled clinical decision support system evaluation framework.MethodsAn artificial intelligence-enabled clinical decision support system evaluation framework consisting of 6 variables was developed. A Delphi process was conducted to develop the measurement instrument items. Cognitive interviews and pretesting were performed to refine the questions. Web-based survey response data were analyzed to remove irrelevant questions from the measurement instrument, to test dimensional structure, and to assess reliability and validity. The interrelationships of relevant variables were tested and verified using path analysis, and a 28-item measurement instrument was developed. Measurement instrument survey responses were collected from 156 respondents.ResultsThe Cronbach α of the measurement instrument was 0.963, and its content validity was 0.943. Values of average variance extracted ranged from 0.582 to 0.756, and values of the heterotrait-monotrait ratio ranged from 0.376 to 0.896. The final model had a good fit (χ262=36.984; P=.08; comparative fit index 0.991; goodness-of-fit index 0.957; root mean square error of approximation 0.052; standardized root mean square residual 0.028). Variables in the final model accounted for 89% of the variance in the user acceptance dimension.ConclusionsUser acceptance is the central dimension of artificial intelligence-enabled clinical decision support system success. Acceptance was directly influenced by perceived ease of use, information quality, service quality, and perceived benefit. Acceptance was also indirectly influenced by system quality and information quality through perceived ease of use. User acceptance and perceived benefit were interrelated.

Project description:BackgroundAssistive automatic seizure detection can empower human annotators to shorten patient monitoring data review times. We present a proof-of-concept for a seizure detection system that is sensitive, automated, patient-specific, and tunable to maximise sensitivity while minimizing human annotation times. The system uses custom data preparation methods, deep learning analytics and electroencephalography (EEG) data.MethodsScalp EEG data of 365 patients containing 171,745 s ictal and 2,185,864 s interictal samples obtained from clinical monitoring systems were analysed as part of a crowdsourced artificial intelligence (AI) challenge. Participants were tasked to develop an ictal/interictal classifier with high sensitivity and low false alarm rates. We built a challenge platform that prevented participants from downloading or directly accessing the data while allowing crowdsourced model development.FindingsThe automatic detection system achieved tunable sensitivities between 75.00% and 91.60% allowing a reduction in the amount of raw EEG data to be reviewed by a human annotator by factors between 142x, and 22x respectively. The algorithm enables instantaneous reviewer-managed optimization of the balance between sensitivity and the amount of raw EEG data to be reviewed.InterpretationThis study demonstrates the utility of deep learning for patient-specific seizure detection in EEG data. Furthermore, deep learning in combination with a human reviewer can provide the basis for an assistive data labelling system lowering the time of manual review while maintaining human expert annotation performance.FundingIBM employed all IBM Research authors. Temple University employed all Temple University authors. The Icahn School of Medicine at Mount Sinai employed Eren Ahsen. The corresponding authors Stefan Harrer and Gustavo Stolovitzky declare that they had full access to all the data in the study and that they had final responsibility for the decision to submit for publication.

Project description:BackgroundDespite the emerging application of clinical decision support systems (CDSS) in pregnancy care and the proliferation of artificial intelligence (AI) over the last decade, it remains understudied regarding the role of AI in CDSS specialized for pregnancy care.ObjectiveTo identify and synthesize AI-augmented CDSS in pregnancy care, CDSS functionality, AI methodologies, and clinical implementation, we reported a systematic review based on empirical studies that examined AI-augmented CDSS in pregnancy care.MethodsWe retrieved studies that examined AI-augmented CDSS in pregnancy care using database queries involved with titles, abstracts, keywords, and MeSH (Medical Subject Headings) terms. Bibliographic records from their inception to 2022 were retrieved from PubMed/MEDLINE (n=206), Embase (n=101), and ACM Digital Library (n=377), followed by eligibility screening and literature review. The eligibility criteria include empirical studies that (1) developed or tested AI methods, (2) developed or tested CDSS or CDSS components, and (3) focused on pregnancy care. Data of studies used for review and appraisal include title, abstract, keywords, MeSH terms, full text, and supplements. Publications with ancillary information or overlapping outcomes were synthesized as one single study. Reviewers independently reviewed and assessed the quality of selected studies.ResultsWe identified 30 distinct studies of 684 studies from their inception to 2022. Topics of clinical applications covered AI-augmented CDSS from prenatal, early pregnancy, obstetric care, and postpartum care. Topics of CDSS functions include diagnostic support, clinical prediction, therapeutics recommendation, and knowledge base.ConclusionsOur review acknowledged recent advances in CDSS studies including early diagnosis of prenatal abnormalities, cost-effective surveillance, prenatal ultrasound support, and ontology development. To recommend future directions, we also noted key gaps from existing studies, including (1) decision support in current childbirth deliveries without using observational data from consequential fetal or maternal outcomes in future pregnancies; (2) scarcity of studies in identifying several high-profile biases from CDSS, including social determinants of health highlighted by the American College of Obstetricians and Gynecologists; and (3) chasm between internally validated CDSS models, external validity, and clinical implementation.

Dataset Information

A framework for evaluating clinical artificial intelligence systems without ground-truth annotations.

Publications

A framework for evaluating clinical artificial intelligence systems without ground-truth annotations.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets