Algorithmic Fairness and Bias Mitigation for Clinical Machine Learning: A New Utility for Deep Reinforcement Learning

As machine learning-based models continue to be developed for healthcare applications, greater effort is needed in ensuring that these technologies do not reflect or exacerbate any unwanted or discriminatory biases that may be present in the data. In this study, we introduce a reinforcement learning framework capable of mitigating biases that may have been acquired during data collection. In particular, we evaluated our model for the task of rapidly predicting COVID-19 for patients presenting to hospital emergency departments, and aimed to mitigate any site-specific (hospital) and ethnicity-based biases present in the data. Using a specialized reward function and training procedure, we show that our method achieves clinically-effective screening performances, while significantly improving outcome fairness compared to current benchmarks and state-of-the-art machine learning methods. We performed external validation across three independent hospitals, and additionally tested our method on a patient ICU discharge status task, demonstrating model generalizability.

Machine Learning Generalizability Across Healthcare Settings: Insights from multi-site COVID-19 screening

As patient health information is highly regulated due to privacy concerns, the majority of machine learning (ML)-based healthcare studies are unable to test on external patient cohorts, resulting in a gap between locally reported model performance and cross-site generalizability. Different approaches have been introduced for developing models across multiple clinical sites, however no studies have compared methods for translating ready-made models for adoption in new settings. We introduce three methods to do this - (1) applying a ready-made model "as-is"; (2) readjusting the decision threshold on the output of a ready-made model using site-specific data; and (3) finetuning a ready-made model using site-specific data via transfer learning. Using a case study of COVID-19 diagnosis across four NHS Hospital Trusts, we show that all methods achieve clinically-effective performances (NPV >0.959), with transfer learning achieving the best results (mean AUROCs between 0.870-0.925). Our models demonstrate that site-specific customization improves predictive performance when compared to other ready-made approaches.

Algorithmic Fairness and Bias Mitigation for Clinical Machine Learning: Insights from Rapid COVID-19 Diagnosis by Adversarial Learning

Machine learning is becoming increasingly prominent in healthcare. Although its benefits are clear, growing attention is being given to how machine learning may exacerbate existing biases and disparities. In this study, we introduce an adversarial training framework that is capable of mitigating biases that may have been acquired through data collection or magnified during model development. For example, if one class is over-presented or errors/inconsistencies in practice are reflected in the training data, then a model can be biased by these. To evaluate our adversarial training framework, we used the statistical definition of equalized odds. We evaluated our model for the task of rapidly predicting COVID-19 for patients presenting to hospital emergency departments, and aimed to mitigate regional (hospital) and ethnic biases present. We trained our framework on a large, real-world COVID-19 dataset and demonstrated that adversarial training demonstrably improves outcome fairness (with respect to equalized odds), while still achieving clinically-effective screening performances (NPV>0.98). We compared our method to the benchmark set by related previous work, and performed prospective and external validation on four independent hospital cohorts. Our method can be generalized to any outcomes, models, and definitions of fairness.

Real-world evaluation of AI driven COVID-19 triage for emergency admissions: External validation & operational assessment of lab-free and high-throughput screening solutions

BackgroundUncertainty in patients COVID-19 status contributes to treatment delays, nosocomial transmission, and operational pressures in hospitals. However, typical turnaround times for batch-processed laboratory PCR tests remain 12-24h. Although rapid antigen lateral flow testing (LFD) has been widely adopted in UK emergency care settings, sensitivity is limited. We recently demonstrated that AI-driven triage (CURIAL-1.0) allows high-throughput COVID-19 screening using clinical data routinely available within 1h of arrival to hospital. Here we aimed to determine operational and safety improvements over standard-care, performing external/prospective evaluation across four NHS trusts with updated algorithms optimised for generalisability and speed, and deploying a novel lab-free screening pathway in a UK emergency department. MethodsWe rationalised predictors in CURIAL-1.0 to optimise separately for generalisability and speed, developing CURIAL-Lab with vital signs and routine laboratory blood predictors (FBC, U&E, LFT, CRP) and CURIAL-Rapide with vital signs and FBC alone. Models were calibrated during training to 90% sensitivity and validated externally for unscheduled admissions to Portsmouth University Hospitals, University Hospitals Birmingham and Bedfordshire Hospitals NHS trusts, and prospectively during the second-wave of the UK COVID-19 epidemic at Oxford University Hospitals (OUH). Predictions were generated using first-performed blood tests and vital signs and compared against confirmatory viral nucleic acid testing. Next, we retrospectively evaluated a novel clinical pathway triaging patients to COVID-19-suspected clinical areas where either model prediction or LFD results were positive, comparing sensitivity and NPV with LFD results alone. Lastly, we deployed CURIAL-Rapide alongside an approved point-of-care FBC analyser (OLO; SightDiagnostics, Israel) to provide lab-free COVID-19 screening in the John Radcliffe Hospitals Emergency Department (Oxford, UK), as trust-approved service improvement. Our primary improvement outcome was time-to-result availability; secondary outcomes were sensitivity, specificity, PPV, and NPV assessed against a PCR reference standard. We compared CURIAL-Rapides performance with clinician triage and LFD results within standard-care. Results72,223 patients met eligibility criteria across external and prospective validation sites. Model performance was consistent across trusts (CURIAL-Lab: AUROCs range 0.858-0.881; CURIAL-Rapide 0.836-0.854), with highest sensitivity achieved at Portsmouth University Hospitals (CURIAL-Lab:84.1% [95% Wilsons score CIs 82.5-85.7]; CURIAL-Rapide:83.5% [81.8 - 85.1]) at specificities of 71.3% (95% Wilsons score CIs: 70.9 - 71.8) and 63.6% (63.1 - 64.1). For 3,207 patients receiving LFD-triage within routine care for OUH admissions between December 23, 2021 and March 6, 2021, a combined clinical pathway increased sensitivity from 56.9% for LFDs alone (95% CI 51.7-62.0) to 88.2% with CURIAL-Rapide (84.4-91.1; AUROC 0.919) and 85.6% with CURIAL-Lab (81.6-88.9; AUROC 0.925). 520 patients were prospectively enrolled for point-of-care FBC analysis between February 18, 2021 and May 10, 2021, of whom 436 received confirmatory PCR testing within routine care and 10 (2.3%) tested positive. Median time from patient arrival to availability of CURIAL-Rapide result was 45:00 min (32-64), 16 minutes (26.3%) sooner than LFD results (61:00 min, 37-99; log-rank p<0.0001), and 6:52 h (90.2%) sooner than PCR results (7:37 h, 6:05-15:39; p<0.0001). Sensitivity and specificity of CURIAL-Rapide were 87.5% (52.9-97.8) and 85.4% (81.3-88.7), therefore achieving high NPV (99.7%, 98.2-99.9). CURIAL-Rapide correctly excluded COVID-19 for 58.5% of negative patients who were triaged by a clinician to COVID-19-suspected (amber) areas. ImpactCURIAL-Lab & CURIAL-Rapide are generalisable, high-throughput screening tests for COVID-19, rapidly excluding the illness with higher NPV than LFDs. CURIAL-Rapide can be used in combination with near-patient FBC analysis for rapid, lab-free screening, and may reduce the number of COVID-19-negative patients triaged to enhanced precautions ( amber) clinical areas.

Artificial intelligence driven assessment of routinely collected healthcare data is an effective screening test for COVID-19 in patients presenting to hospital

BackgroundRapid identification of COVID-19 is important for delivering care expediently and maintaining infection control. The early clinical course of SARS-CoV-2 infection can be difficult to distinguish from other undifferentiated medical presentations to hospital, however for operational reasons SARS-CoV-2 PCR testing can take up to 48 hours. Artificial Intelligence (AI) methods, trained using routinely collected clinical data, may allow front-door screening for COVID-19 within the first hour of presentation. MethodsDemographic, routine and prior clinical data were extracted for 170,510 sequential presentations to emergency and acute medical departments at a large UK teaching hospital group. We applied multivariate logistic regression, random forests and extreme gradient boosted trees to distinguish emergency department (ED) presentations and admissions due to COVID-19 from pre-pandemic controls. We performed stepwise addition of clinical feature sets and assessed performance using stratified 10-fold cross validation. Models were calibrated during training to achieve sensitivities of 70, 80 and 90% for identifying patients with COVID-19. To simulate real-world performance at different stages of an epidemic, we generated test sets with varying prevalences of COVID-19 and assessed predictive values. We prospectively validated our models for all patients presenting or admitted to our hospital group between 20th April and 6th May 2020, comparing model predictions to PCR test results. ResultsPresentation laboratory blood tests, point of care blood gas, and vital signs measurements for 115,394 emergency presentations and 72,310 admissions were analysed. Presentation laboratory tests and vital signs were most predictive of COVID-19 (maximum area under ROC curve [AUROC] 0.904 and 0.823, respectively). Sequential addition of informative variables improved model performance to AUROC 0.942. We developed two early-detection models to identify COVID-19, achieving sensitivities and specificities of 77.4% and 95.7% for our ED model amongst patients attending hospital, and 77.4% and 94.8% for our Admissions model amongst patients being admitted. Both models offer high negative predictive values (>99%) across a range of prevalences (<5%). In a two-week prospective validation period, our ED and Admissions models demonstrated 92.3% and 92.5% accuracy (AUROC 0.881 and 0.871 respectively) for all patients presenting or admitted to a large UK teaching hospital group. A sensitivity analysis to account for uncertainty in negative PCR results improves apparent accuracy (95.1% and 94.1%) and NPV (99.0% and 98.5%). Three laboratory blood markers, Eosinophils, Basophils, and C-Reactive Protein, alongside Calcium measured on blood-gas, and presentation Oxygen requirement were the most informative variables in our models. ConclusionArtificial intelligence techniques perform effectively as a screening test for COVID-19 in emergency departments and hospital admission units. Our models support rapid exclusion of the illness using routinely collected and readily available clinical measurements, guiding streaming of patients during the early phase of admission. BriefThe early clinical course of SARS-CoV-2 infection can be difficult to distinguish from other undifferentiated medical presentations to hospital, however viral specific real-time polymerase chain reaction (RT-PCR) testing has limited sensitivity and can take up to 48 hours for operational reasons. In this study, we develop two early-detection models to identify COVID-19 using routinely collected data typically available within one hour (laboratory tests, blood gas and vital signs) during 115,394 emergency presentations and 72,310 admissions to hospital. Our emergency department (ED) model achieved 77.4% sensitivity and 95.7% specificity (AUROC 0.939) for COVID-19 amongst all patients attending hospital, and Admissions model achieved 77.4% sensitivity and 94.8% specificity (AUROC 0.940) for the subset admitted to hospital. Both models achieve high negative predictive values (>99%) across a range of prevalences (<5%), facilitating rapid exclusion during triage to guide infection control. We prospectively validated our models across all patients presenting and admitted to a large UK teaching hospital group in a two-week test period, achieving 92.3% (n= 3,326, NPV: 97.6%, AUROC: 0.881) and 92.5% accuracy (n=1,715, NPV: 97.7%, AUROC: 0.871) in comparison to RT-PCR results. Sensitivity analyses to account for uncertainty in negative PCR results improves apparent accuracy (95.1% and 94.1%) and NPV (99.0% and 98.5%). Our artificial intelligence models perform effectively as a screening test for COVID-19 in emergency departments and hospital admission units, offering high impact in settings where rapid testing is unavailable.