Virtual Biopsy by Using Artificial Intelligence-based Multimodal Modeling of Binational Mammography Data.

Background Computational models based on artificial intelligence (AI) are increasingly used to diagnose malignant breast lesions. However, assessment from radiologic images of the specific pathologic lesion subtypes, as detailed in the results of biopsy procedures, remains a challenge. Purpose To develop an AI-based model to identify breast lesion subtypes with mammograms and linked electronic health records labeled with histopathologic information. Materials and Methods In this retrospective study, 26 569 images were collected in 9234 women who underwent digital mammography to pretrain the algorithms. The training data included individuals who had at least 1 year of clinical and imaging history followed by biopsy-based histopathologic diagnosis from March 2013 to November 2018. A model that combined convolutional neural networks with supervised learning algorithms was independently trained to make breast lesion predictions with data from 2120 women in Israel and 1642 women in the United States. Results were reported using the area under the receiver operating characteristic curve (AUC) with the 95% DeLong approach to estimate CIs. Significance was tested with bootstrapping. Results The Israeli model was validated in 456 women and tested in 441 women (mean age, 51 years ± 11 [SD]). The U.S. model was validated in 350 women and tested in 344 women (mean age, 60 years ± 12). For predicting malignancy in the test sets (consisting of 220 Israeli patient examinations and 126 U.S. patient examinations with ductal carcinoma in situ or invasive cancer), the algorithms obtained an AUC of 0.88 (95% CI: 0.85, 0.91) and 0.80 (95% CI: 0.74, 0.85) for Israeli and U.S. patients, respectively (P = .006). These results may not hold for other cohorts of patients, and generalizability across populations should be further investigated. Conclusion The results offer supporting evidence that artificial intelligence applied to clinical and mammographic images can identify breast lesion subtypes when the data are sufficiently large, which may help assess diagnostic workflow and reduce biopsy sampling errors. Published under a CC BY 4.0 license. Online supplemental material is available for this article.

Artificial Intelligence for Reducing Workload in Breast Cancer Screening with Digital Breast Tomosynthesis.

Background Digital breast tomosynthesis (DBT) has higher diagnostic accuracy than digital mammography, but interpretation time is substantially longer. Artificial intelligence (AI) could improve reading efficiency. Purpose To evaluate the use of AI to reduce workload by filtering out normal DBT screens. Materials and Methods The retrospective study included 13 306 DBT examinations from 9919 women performed between June 2013 and November 2018 from two health care networks. The cohort was split into training, validation, and test sets (3948, 1661, and 4310 women, respectively). A workflow was simulated in which the AI model classified cancer-free examinations that could be dismissed from the screening worklist and used the original radiologists' interpretations on the rest of the worklist examinations. The AI system was also evaluated with a reader study of five breast radiologists reading the DBT mammograms of 205 women. The area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and recall rate were evaluated in both studies. Statistics were computed across 10 000 bootstrap samples to assess 95% CIs, noninferiority, and superiority tests. Results The model was tested on 4310 screened women (mean age, 60 years ± 11 [standard deviation]; 5182 DBT examinations). Compared with the radiologists' performance (417 of 459 detected cancers [90.8%], 477 recalls in 5182 examinations [9.2%]), the use of AI to automatically filter out cases would result in 39.6% less workload, noninferior sensitivity (413 of 459 detected cancers; 90.0%; P = .002), and 25% lower recall rate (358 recalls in 5182 examinations; 6.9%; P = .002). In the reader study, AUC was higher in the standalone AI compared with the mean reader (0.84 vs 0.81; P = .002). Conclusion The artificial intelligence model was able to identify normal digital breast tomosynthesis screening examinations, which decreased the number of examinations that required radiologist interpretation in a simulated clinical workflow. Published under a CC BY 4.0 license. Online supplemental material is available for this article. See also the editorial by Philpotts in this issue.

Predicting Breast Cancer by Applying Deep Learning to Linked Health Records and Mammograms.

Background Computational models on the basis of deep neural networks are increasingly used to analyze health care data. However, the efficacy of traditional computational models in radiology is a matter of debate. Purpose To evaluate the accuracy and efficiency of a combined machine and deep learning approach for early breast cancer detection applied to a linked set of digital mammography images and electronic health records. Materials and Methods In this retrospective study, 52 936 images were collected in 13 234 women who underwent at least one mammogram between 2013 and 2017, and who had health records for at least 1 year before undergoing mammography. The algorithm was trained on 9611 mammograms and health records of women to make two breast cancer predictions: to predict biopsy malignancy and to differentiate normal from abnormal screening examinations. The study estimated the association of features with outcomes by using t test and Fisher exact test. The model comparisons were performed with a 95% confidence interval (CI) or by using the DeLong test. Results The resulting algorithm was validated in 1055 women and tested in 2548 women (mean age, 55 years ± 10 [standard deviation]). In the test set, the algorithm identified 34 of 71 (48%) false-negative findings on mammograms. For the malignancy prediction objective, the algorithm obtained an area under the receiver operating characteristic curve (AUC) of 0.91 (95% CI: 0.89, 0.93), with specificity of 77.3% (95% CI: 69.2%, 85.4%) at a sensitivity of 87%. When trained on clinical data alone, the model performed significantly better than the Gail model (AUC, 0.78 vs 0.54, respectively; P < .004). Conclusion The algorithm, which combined machine-learning and deep-learning approaches, can be applied to assess breast cancer at a level comparable to radiologists and has the potential to substantially reduce missed diagnoses of breast cancer. © RSNA, 2019 Online supplemental material is available for this article.

Changing the approach to treatment choice in epilepsy using big data.

PURPOSE: A UCB-IBM collaboration explored the application of machine learning to large claims databases to construct an algorithm for antiepileptic drug (AED) choice for individual patients. METHODS: Claims data were collected between January 2006 and September 2011 for patients with epilepsy > 16 years of age. A subset of patient claims with a valid index date of AED treatment change (new, add, or switch) were used to train the AED prediction model by retrospectively evaluating an index date treatment for subsequent treatment change. Based on the trained model, a model-predicted AED regimen with the lowest likelihood of treatment change was assigned to each patient in the group of test claims, and outcomes were evaluated to test model validity. RESULTS: The model had 72% area under receiver operator characteristic curve, indicating good predictive power. Patients who were given the model-predicted AED regimen had significantly longer survival rates (time until a treatment change event) and lower expected health resource utilization on average than those who received another treatment. The actual prescribed AED regimen at the index date matched the model-predicted AED regimen in only 13% of cases; there were large discrepancies in the frequency of use of certain AEDs/combinations between model-predicted AED regimens and those actually prescribed. CONCLUSIONS: Chances of treatment success were improved if patients received the model-predicted treatment. Using the model's prediction system may enable personalized, evidence-based epilepsy care, accelerating the match between patients and their ideal therapy, thereby delivering significantly better health outcomes for patients and providing health-care savings by applying resources more efficiently. Our goal will be to strengthen the predictive power of the model by integrating diverse data sets and potentially moving to prospective data collection.

A deep learning framework combining molecular image and protein structural representations identifies candidate drugs for pain.

Artificial intelligence (AI) and deep learning technologies hold promise for identifying effective drugs for human diseases, including pain. Here, we present an interpretable deep-learning-based ligand image- and receptor's three-dimensional (3D)-structure-aware framework to predict compound-protein interactions (LISA-CPI). LISA-CPI integrates an unsupervised deep-learning-based molecular image representation (ImageMol) of ligands and an advanced AlphaFold2-based algorithm (Evoformer). We demonstrated that LISA-CPI achieved ∼20% improvement in the average mean absolute error (MAE) compared to state-of-the-art models on experimental CPIs connecting 104,969 ligands and 33 G-protein-coupled receptors (GPCRs). Using LISA-CPI, we prioritized potential repurposable drugs (e.g., methylergometrine) and identified candidate gut-microbiota-derived metabolites (e.g., citicoline) for potential treatment of pain via specifically targeting human GPCRs. In summary, we presented that the integration of molecular image and protein 3D structural representations using a deep learning framework offers a powerful computational drug discovery tool for treating pain and other complex diseases if broadly applied.

Comparing the efficacy of anti-seizure medications using matched cohorts on a large insurance claims database.

Epilepsy is a severe chronic neurological disease affecting 60 million people worldwide. Primary treatment is with anti-seizure medicines (ASMs), but many patients continue to experience seizures. We used retrospective insurance claims data on 280,587 patients with uncontrolled epilepsy (UE), defined as status epilepticus, need for a rescue medicine, or admission or emergency visit for an epilepsy code. We conducted a computational risk ratio analysis between pairs of ASMs using a causal inference method, in order to match 1034 clinical factors and simulate randomization. Data was extracted from the MarketScan insurance claims Research Database records from 2011 to 2015. The cohort consisted of individuals over 18 years old with a diagnosis of epilepsy who took one of eight ASMs and had more than a year of history prior to the filling of the drug prescription. Seven ASM exposures were analyzed: topiramate, phenytoin, levetiracetam, gabapentin, lamotrigine, valproate, and carbamazepine or oxcarbazepine (treated as the same exposure). We calculated the risk ratio of UE between pairs of ASM after controlling for bias with inverse propensity weighting applied to 1034 factors, such as demographics, confounding illnesses, non-epileptic conditions treated by ASMs, etc. All ASMs exhibited a significant reduction in the prevalence of UE, but three drugs showed pair-wise differences compared to other ASMs. Topiramate consistently was associated with a lower risk of UE, with a mean risk ratio range of 0.68-0.93 (average 0.82, CI: 0.56-1.08). Phenytoin and levetiracetam were consistently associated with a higher risk of UE with mean risk ratio ranges of 1.11 to 1.47 (average 1.13, CI 0.98-1.65) and 1.15 to 1.43 (average 1.2, CI 0.72-1.69), respectively. Large-scale retrospective insurance claims data - combined with causal inference analysis - provides an opportunity to compare the effect of treatments in real-world data in populations 1,000-fold larger than those in typical randomized trials. Our causal analysis identified the clinically unexpected finding of topiramate as being associated with a lower risk of UE; and phenytoin and levetiracetam as associated with a higher risk of UE (compared to other studied drugs, not to baseline). However, we note that our data set for this study only used insurance claims events, which does not comprise actual seizure frequencies, nor a clear picture of side effects. Our results do not advocate for any change in practice but demonstrate that conclusions from large databases may differ from and supplement those of randomized trials and clinical practice and therefore may guide further investigation.

From Pixels to Diagnosis: Algorithmic Analysis of Clinical Oral Photos for Early Detection of Oral Squamous Cell Carcinoma.

Oral squamous cell carcinoma (OSCC) accounts for more than 90% of oral malignancies. Despite numerous advancements in understanding its biology, the mean five-year survival rate of OSCC is still very poor at about 50%, with even lower rates when the disease is detected at later stages. We investigate the use of clinical photographic images taken by common smartphones for the automated detection of OSCC cases and for the identification of suspicious cases mimicking cancer that require an urgent biopsy. We perform a retrospective study on a cohort of 1470 patients drawn from both hospital records and online academic sources. We examine various deep learning methods for the early detection of OSCC cases as well as for the detection of suspicious cases. Our results demonstrate the efficacy of these methods in both tasks, providing a comprehensive understanding of the patient's condition. When evaluated on holdout data, the model to predict OSCC achieved an AUC of 0.96 (CI: 0.91, 0.98), with a sensitivity of 0.91 and specificity of 0.81. When the data are stratified based on lesion location, we find that our models can provide enhanced accuracy (AUC 1.00) in differentiating specific groups of patients that have lesions in the lingual mucosa, floor of mouth, or posterior tongue. These results underscore the potential of leveraging clinical photos for the timely and accurate identification of OSCC.

Artificial intelligence and open science in discovery of disease-modifying medicines for Alzheimer's disease.

The high failure rate of clinical trials in Alzheimer's disease (AD) and AD-related dementia (ADRD) is due to a lack of understanding of the pathophysiology of disease, and this deficit may be addressed by applying artificial intelligence (AI) to "big data" to rapidly and effectively expand therapeutic development efforts. Recent accelerations in computing power and availability of big data, including electronic health records and multi-omics profiles, have converged to provide opportunities for scientific discovery and treatment development. Here, we review the potential utility of applying AI approaches to big data for discovery of disease-modifying medicines for AD/ADRD. We illustrate how AI tools can be applied to the AD/ADRD drug development pipeline through collaborative efforts among neurologists, gerontologists, geneticists, pharmacologists, medicinal chemists, and computational scientists. AI and open data science expedite drug discovery and development of disease-modifying therapeutics for AD/ADRD and other neurodegenerative diseases.

Nurturing next generation physicians: A new Israeli healthtech fellowship.

The Israeli Society for HealthTech aims at advancing the integration of innovation and healthcare entrepreneurship into medical practice and across traditional health professions, to benefit patients and improve quality of care. In 2021, the Society launched the first fellowship for board certified physicians in HealthTech. This backstory discusses the motivation of launching the program and reviews the design of the fellowship, including curriculum, the expertise of the lecturers, and initial tangible results of the program.

Impact of the COVID-19 Pandemic on Clinical Findings in Medical Imaging Exams in a Nationwide Israeli Health Organization: Observational Study.

BACKGROUND: The outbreak of the COVID-19 pandemic had a major effect on the consumption of health care services. Changes in the use of routine diagnostic exams, increased incidences of postacute COVID-19 syndrome (PCS), and other pandemic-related factors may have influenced detected clinical conditions. OBJECTIVE: This study aimed to analyze the impact of COVID-19 on the use of outpatient medical imaging services and clinical findings therein, specifically focusing on the time period after the launch of the Israeli COVID-19 vaccination campaign. In addition, the study tested whether the observed gains in abnormal findings may be linked to PCS or COVID-19 vaccination. METHODS: Our data set included 572,480 ambulatory medical imaging patients in a national health organization from January 1, 2019, to August 31, 2021. We compared different measures of medical imaging utilization and clinical findings therein before and after the surge of the pandemic to identify significant changes. We also inspected the changes in the rate of abnormal findings during the pandemic after adjusting for changes in medical imaging utilization. Finally, for imaging classes that showed increased rates of abnormal findings, we measured the causal associations between SARS-CoV-2 infection, COVID-19-related hospitalization (indicative of COVID-19 complications), and COVID-19 vaccination and future risk for abnormal findings. To adjust for a multitude of confounding factors, we used causal inference methodologies. RESULTS: After the initial drop in the utilization of routine medical imaging due to the first COVID-19 wave, the number of these exams has increased but with lower proportions of older patients, patients with comorbidities, women, and vaccine-hesitant patients. Furthermore, we observed significant gains in the rate of abnormal findings, specifically in musculoskeletal magnetic resonance (MR-MSK) and brain computed tomography (CT-brain) exams. These results also persisted after adjusting for the changes in medical imaging utilization. Demonstrated causal associations included the following: SARS-CoV-2 infection increasing the risk for an abnormal finding in a CT-brain exam (odds ratio [OR] 1.4, 95% CI 1.1-1.7) and COVID-19-related hospitalization increasing the risk for abnormal findings in an MR-MSK exam (OR 3.1, 95% CI 1.9-5.3). CONCLUSIONS: COVID-19 impacted the use of ambulatory imaging exams, with greater avoidance among patients at higher risk for COVID-19 complications: older patients, patients with comorbidities, and nonvaccinated patients. Causal analysis results imply that PCS may have contributed to the observed gains in abnormal findings in MR-MSK and CT-brain exams.

A historical perspective of biomedical explainable AI research.

The black-box nature of most artificial intelligence (AI) models encourages the development of explainability methods to engender trust into the AI decision-making process. Such methods can be broadly categorized into two main types: post hoc explanations and inherently interpretable algorithms. We aimed at analyzing the possible associations between COVID-19 and the push of explainable AI (XAI) to the forefront of biomedical research. We automatically extracted from the PubMed database biomedical XAI studies related to concepts of causality or explainability and manually labeled 1,603 papers with respect to XAI categories. To compare the trends pre- and post-COVID-19, we fit a change point detection model and evaluated significant changes in publication rates. We show that the advent of COVID-19 in the beginning of 2020 could be the driving factor behind an increased focus concerning XAI, playing a crucial role in accelerating an already evolving trend. Finally, we present a discussion with future societal use and impact of XAI technologies and potential future directions for those who pursue fostering clinical trust with interpretable machine learning models.

Predicting response to antiretroviral treatment by machine learning: the EuResist project.

For a long time, the clinical management of antiretroviral drug resistance was based on sequence analysis of the HIV genome followed by estimating drug susceptibility from the mutational pattern that was detected. The large number of anti-HIV drugs and HIV drug resistance mutations has prompted the development of computer-aided genotype interpretation systems, typically comprising rules handcrafted by experts via careful examination of in vitro and in vivo resistance data. More recently, machine learning approaches have been applied to establish data-driven engines able to indicate the most effective treatments for any patient and virus combination. Systems of this kind, currently including the Resistance Response Database Initiative and the EuResist engine, must learn from the large data sets of patient histories and can provide an objective and accurate estimate of the virological response to different antiretroviral regimens. The EuResist engine was developed by a European consortium of HIV and bioinformatics experts and compares favorably with the most commonly used genotype interpretation systems and HIV drug resistance experts. Next-generation treatment response prediction engines may valuably assist the HIV specialist in the challenging task of establishing effective regimens for patients harboring drug-resistant virus strains. The extensive collection and accurate processing of increasingly large patient data sets are eagerly awaited to further train and translate these systems from prototype engines into real-life treatment decision support tools.

Efficacy of antiretroviral therapy switch in HIV-infected patients: a 10-year analysis of the EuResist Cohort.

INTRODUCTION: Highly active antiretroviral therapy (HAART) has been shown to be effective in many recent trials. However, there is limited data on time trends of HAART efficacy after treatment change. METHODS: Data from different European cohorts were compiled within the EuResist Project. The efficacy of HAART defined by suppression of viral replication at 24 weeks after therapy switch was analyzed considering previous treatment modifications from 1999 to 2008. RESULTS: Altogether, 12,323 treatment change episodes in 7,342 patients were included in the analysis. In 1999, HAART after treatment switch was effective in 38.0% of the patients who had previously undergone 1-5 therapies. This figure rose to 85.0% in 2008. In patients with more than 5 previous therapies, efficacy rose from 23.9 to 76.2% in the same time period. In patients with detectable viral load at therapy switch, the efficacy rose from 23.3 to 66.7% with 1-5 previous treatments and from 14.4 to 55.6% with more than 5 previous treatments. CONCLUSION: The results of this large cohort show that the outcome of HAART switch has improved considerably over the last years. This result was particularly observed in the context after viral rebound. Thus, changing HAART is no longer associated with a high risk of treatment failure.

Quantification of tumor heterogeneity: from data acquisition to metric generation.

Tumors are unique and complex ecosystems, in which heterogeneous cell subpopulations with variable molecular profiles, aggressiveness, and proliferation potential coexist and interact. Understanding how heterogeneity influences tumor progression has important clinical implications for improving diagnosis, prognosis, and treatment response prediction. Several recent innovations in data acquisition methods and computational metrics have enabled the quantification of spatiotemporal heterogeneity across different scales of tumor organization. Here, we summarize the most promising efforts from a common experimental and computational perspective, discussing their advantages, shortcomings, and challenges. With personalized medicine entering a new era of unprecedented opportunities, our vision is that of future workflows integrating across modalities, scales, and dimensions to capture intricate aspects of the tumor ecosystem and to open new avenues for improved patient care.

Multimodal Prediction of Five-Year Breast Cancer Recurrence in Women Who Receive Neoadjuvant Chemotherapy.

In current clinical practice, it is difficult to predict whether a patient receiving neoadjuvant chemotherapy (NAC) for breast cancer is likely to encounter recurrence after treatment and have the cancer recur locally in the breast or in other areas of the body. We explore the use of clinical history, immunohistochemical markers, and multiparametric magnetic resonance imaging (DCE, ADC, Dixon) to predict the risk of post-treatment recurrence within five years. We performed a retrospective study on a cohort of 1738 patients from Institut Curie and analyzed the data using classical machine learning, image processing, and deep learning. Our results demonstrate the ability to predict recurrence prior to NAC treatment initiation using each modality alone, and the possible improvement achieved by combining the modalities. When evaluated on holdout data, the multimodal model achieved an AUC of 0.75 (CI: 0.70, 0.80) and 0.57 specificity at 0.90 sensitivity. We then stratified the data based on known prognostic biomarkers. We found that our models can provide accurate recurrence predictions (AUC > 0.89) for specific groups of women under 50 years old with poor prognoses. A version of our method won second place at the BMMR2 Challenge, with a very small margin from being first, and was a standout from the other challenge entries.

A causal inference approach for estimating effects of non-pharmaceutical interventions during Covid-19 pandemic.

In response to the outbreak of the coronavirus disease 2019 (Covid-19), governments worldwide have introduced multiple restriction policies, known as non-pharmaceutical interventions (NPIs). However, the relative impact of control measures and the long-term causal contribution of each NPI are still a topic of debate. We present a method to rigorously study the effectiveness of interventions on the rate of the time-varying reproduction number Rt and on human mobility, considered here as a proxy measure of policy adherence and social distancing. We frame our model using a causal inference approach to quantify the impact of five governmental interventions introduced until June 2020 to control the outbreak in 113 countries: confinement, school closure, mask wearing, cultural closure, and work restrictions. Our results indicate that mobility changes are more accurately predicted when compared to reproduction number. All NPIs, except for mask wearing, significantly affected human mobility trends. From these, schools and cultural closure mandates showed the largest effect on social distancing. We also found that closing schools, issuing face mask usage, and work-from-home mandates also caused a persistent reduction on Rt after their initiation, which was not observed with the other social distancing measures. Our results are robust and consistent across different model specifications and can shed more light on the impact of individual NPIs.

Leveraging Comprehensive Health Records for Breast Cancer Risk Prediction: A Binational Assessment.

Breast cancer (BC) risk models based on electronic health records (EHR) can assist physicians in estimating the probability of an individual with certain risk factors to develop BC in the future. In this retrospective study, we used clinical data combined with machine learning tools to assess the utility of a personalized BC risk model on 13,786 Israeli and 1,695 American women who underwent screening mammography in the years 2012-2018 and 2008-2018, respectively. Clinical features were extracted from EHR, personal questionnaires, and past radiologists' reports. Using a set of 1,547 features, the predictive ability for BC within 12 months was measured in both datasets and in sub-cohorts of interest. Our results highlight the improved performance of our model over previous established BC risk models, their ultimate potential for risk-based screening policies on first time patients and novel clinically relevant risk factors that can compensate for the absence of imaging history information.

Emulated Clinical Trials from Longitudinal Real-World Data Efficiently Identify Candidates for Neurological Disease Modification: Examples from Parkinson's Disease.

Real-world healthcare data hold the potential to identify therapeutic solutions for progressive diseases by efficiently pinpointing safe and efficacious repurposing drug candidates. This approach circumvents key early clinical development challenges, particularly relevant for neurological diseases, concordant with the vision of the 21st Century Cures Act. However, to-date, these data have been utilized mainly for confirmatory purposes rather than as drug discovery engines. Here, we demonstrate the usefulness of real-world data in identifying drug repurposing candidates for disease-modifying effects, specifically candidate marketed drugs that exhibit beneficial effects on Parkinson's disease (PD) progression. We performed an observational study in cohorts of ascertained PD patients extracted from two large medical databases, Explorys SuperMart (N = 88,867) and IBM MarketScan Research Databases (N = 106,395); and applied two conceptually different, well-established causal inference methods to estimate the effect of hundreds of drugs on delaying dementia onset as a proxy for slowing PD progression. Using this approach, we identified two drugs that manifested significant beneficial effects on PD progression in both datasets: rasagiline, narrowly indicated for PD motor symptoms; and zolpidem, a psycholeptic. Each confers its effects through distinct mechanisms, which we explored via a comparison of estimated effects within the drug classification ontology. We conclude that analysis of observational healthcare data, emulating otherwise costly, large, and lengthy clinical trials, can highlight promising repurposing candidates, to be validated in prospective registration trials, beneficial against common, late-onset progressive diseases for which disease-modifying therapeutic solutions are scarce.

Predictive and Causal Analysis of No-Shows for Medical Exams During COVID-19: A Case Study of Breast Imaging in a Nationwide Israeli Health Organization.

"No-shows", defined as missed appointments or late cancellations, is a central problem in healthcare systems. It has appeared to intensify during the COVID-19 pandemic and the nonpharmaceutical interventions, such as closures, taken to slow its spread. No-shows interfere with patients' continuous care, lead to inefficient utilization of medical resources, and increase healthcare costs. We present a comprehensive analysis of no-shows for breast imaging appointments made during 2020 in a large medical network in Israel. We applied advanced machine learning methods to provide insights into novel and known predictors. Additionally, we employed causal inference methodology to infer the effect of closures on no-shows, after accounting for confounding biases, and demonstrate the superiority of adversarial balancing over inverse probability weighting in correcting these biases. Our results imply that a patient's perceived risk of cancer and the COVID-19 time-based factors are major predictors. Further, we reveal that closures impact patients over 60, but not patients undergoing advanced diagnostic examinations.