Advancements in Large Vision-Language Models (LVLMs) have demonstrated impressive performance in image-conditioned text generation; however, hallucinated outputs-text that misaligns with the visual input-pose a major barrier to their use in safety-critical applications.We introduce CONFLVLM, a conformal-prediction-based framework that achieves finite-sample distribution-free statistical guarantees to the factuality of LVLM output.Taking each generated detail as a hypothesis, CONFLVLM statistically tests factuality via efficient heuristic uncertainty measures to filter out unreliable claims.We conduct extensive experiments covering three representative application domains: general scene understanding, medical radiology report generation, and document understanding.Remarkably, CON-FLVLM reduces the error rate of claims generated by LLaVa-1.5 for scene descriptions from 87.8% to 10.0% by filtering out erroneous claims with a 95.3% true positive rate.Our results further show that CONFLVLM is highly flexible, and can be applied to any black-box LVLMs paired with any uncertainty measure for any image-conditioned free-form text generation task while providing a rigorous guarantee on controlling hallucination risk.

OBJECTIVES: To determine the extent to which current large language models (LLMs) can serve as substitutes for traditional machine learning (ML) as clinical predictors using data from electronic health records (EHRs), we investigated various factors that can impact their adoption, including overall performance, calibration, fairness, and resilience to privacy protections that reduce data fidelity. MATERIALS AND METHODS: We evaluated GPT-3.5, GPT-4, and traditional ML (as gradient-boosting trees) on clinical prediction tasks in EHR data from Vanderbilt University Medical Center (VUMC) and MIMIC IV. We measured predictive performance with area under the receiver operating characteristic (AUROC) and model calibration using Brier Score. To evaluate the impact of data privacy protections, we assessed AUROC when demographic variables are generalized. We evaluated algorithmic fairness using equalized odds and statistical parity across race, sex, and age of patients. We also considered the impact of using in-context learning by incorporating labeled examples within the prompt. RESULTS: Traditional ML [AUROC: 0.847, 0.894 (VUMC, MIMIC)] substantially outperformed GPT-3.5 (AUROC: 0.537, 0.517) and GPT-4 (AUROC: 0.629, 0.602) (with and without in-context learning) in predictive performance and output probability calibration [Brier Score (ML vs GPT-3.5 vs GPT-4): 0.134 vs 0.384 vs 0.251, 0.042 vs 0.06 vs 0.219)]. DISCUSSION: Traditional ML is more robust than GPT-3.5 and GPT-4 in generalizing demographic information to protect privacy. GPT-4 is the fairest model according to our selected metrics but at the cost of poor model performance. CONCLUSION: These findings suggest that non-fine-tuned LLMs are less effective and robust than locally trained ML for clinical prediction tasks, but they are improving across releases.

OBJECTIVE: Diagnosis codes documented in electronic health records (EHR) are often relied upon to clinically phenotype patients for biomedical research. However, these diagnoses can be incomplete and inaccurate, leading to false negatives when searching for patients with phenotypes of interest. This study aims to determine whether PheMAP, a comprehensive knowledgebase integrating multiple clinical terminologies beyond diagnosis to capture phenotypes, can effectively identify patients lacking relevant EHR diagnosis codes. MATERIALS AND METHODS: We investigated a collection of 3.5 million patient records from Vanderbilt University Medical Center's EHR and focused on 4 well-studied phenotypes: (1) type 2 diabetes mellitus (T2DM), (2) dementia, (3) prostate cancer, and (4) sensorineural hearing loss. We applied PheMAP to match structured concepts in patient records and calculated a phenotype risk score (PheScore) to indicate patient-phenotype similarity. Patients meeting predefined PheScore criteria but lacking diagnosis codes were identified. Clinically knowledgeable experts adjudicated randomly selected patients per phenotype as Positive, Possibly Positive, or Negative. RESULTS: Our approach indicated that 5.3% of patients lacked a diagnosis for T2DM, 4.5% for dementia, 2.2% for prostate cancer, and 0.2% for sensorineural hearing loss. The expert review indicated 100% precision (for Possibly Positive or Positive cases) for dementia and sensorineural hearing loss, and 90.0% and 85.0% precision for T2DM and prostate cancer, respectively. Excluding Possibly Positive cases, the precision for T2DM and prostate cancer was 88.9% and 81.3%, respectively. CONCLUSIONS: Leveraging clinical terminologies incorporated by PheMAP can effectively identify patients with phenotypes who lack EHR diagnosis codes, thereby enhancing phenotyping quality and related research reliability.

Language models (LMs) are known to suffer from hallucinations and misinformation.Retrieval augmented generation (RAG) that retrieves verifiable information from an external knowledge corpus to complement the parametric knowledge in LMs provides a tangible solution to these problems.However, the generation quality of RAG is highly dependent on the relevance between a user's query and the retrieved documents.Inaccurate responses may be generated when the query is outside of the scope of knowledge represented in the external knowledge corpus or if the information in the corpus is out-of-date.In this work, we establish a statistical framework that assesses how well a query can be answered by an RAG system by capturing the relevance of knowledge.We introduce an online testing procedure that employs goodness-of-fit (GoF) tests to inspect the relevance of each user query to detect out-of-knowledge queries with low knowledge relevance.Additionally, we develop an offline testing framework that examines a collection of user queries, aiming to detect significant shifts in the query distribution which indicates the knowledge corpus is no longer sufficiently capable of supporting the interests of the users.We demonstrate the capabilities of these strategies through a systematic evaluation on eight question-answering (QA) datasets, the results of which indicate that the new testing framework is an efficient solution to enhance the reliability of existing RAG systems.

OBJECTIVES: Racial disparities in kidney transplant access and posttransplant outcomes exist between non-Hispanic Black (NHB) and non-Hispanic White (NHW) patients in the United States, with the site of care being a key contributor. Using multi-site data to examine the effect of site of care on racial disparities, the key challenge is the dilemma in sharing patient-level data due to regulations for protecting patients' privacy. MATERIALS AND METHODS: We developed a federated learning framework, named dGEM-disparity (decentralized algorithm for Generalized linear mixed Effect Model for disparity quantification). Consisting of 2 modules, dGEM-disparity first provides accurately estimated common effects and calibrated hospital-specific effects by requiring only aggregated data from each center and then adopts a counterfactual modeling approach to assess whether the graft failure rates differ if NHB patients had been admitted at transplant centers in the same distribution as NHW patients were admitted. RESULTS: Utilizing United States Renal Data System data from 39 043 adult patients across 73 transplant centers over 10 years, we found that if NHB patients had followed the distribution of NHW patients in admissions, there would be 38 fewer deaths or graft failures per 10 000 NHB patients (95% CI, 35-40) within 1 year of receiving a kidney transplant on average. DISCUSSION: The proposed framework facilitates efficient collaborations in clinical research networks. Additionally, the framework, by using counterfactual modeling to calculate the event rate, allows us to investigate contributions to racial disparities that may occur at the level of site of care. CONCLUSIONS: Our framework is broadly applicable to other decentralized datasets and disparities research related to differential access to care. Ultimately, our proposed framework will advance equity in human health by identifying and addressing hospital-level racial disparities.

OBJECTIVES: Phenotyping is a core task in observational health research utilizing electronic health records (EHRs). Developing an accurate algorithm demands substantial input from domain experts, involving extensive literature review and evidence synthesis. This burdensome process limits scalability and delays knowledge discovery. We investigate the potential for leveraging large language models (LLMs) to enhance the efficiency of EHR phenotyping by generating high-quality algorithm drafts. MATERIALS AND METHODS: We prompted four LLMs-GPT-4 and GPT-3.5 of ChatGPT, Claude 2, and Bard-in October 2023, asking them to generate executable phenotyping algorithms in the form of SQL queries adhering to a common data model (CDM) for three phenotypes (ie, type 2 diabetes mellitus, dementia, and hypothyroidism). Three phenotyping experts evaluated the returned algorithms across several critical metrics. We further implemented the top-rated algorithms and compared them against clinician-validated phenotyping algorithms from the Electronic Medical Records and Genomics (eMERGE) network. RESULTS: GPT-4 and GPT-3.5 exhibited significantly higher overall expert evaluation scores in instruction following, algorithmic logic, and SQL executability, when compared to Claude 2 and Bard. Although GPT-4 and GPT-3.5 effectively identified relevant clinical concepts, they exhibited immature capability in organizing phenotyping criteria with the proper logic, leading to phenotyping algorithms that were either excessively restrictive (with low recall) or overly broad (with low positive predictive values). CONCLUSION: GPT versions 3.5 and 4 are capable of drafting phenotyping algorithms by identifying relevant clinical criteria aligned with a CDM. However, expertise in informatics and clinical experience is still required to assess and further refine generated algorithms.

Text-attributed graphs (TAGs) are prevalent in various real-world scenarios, where each node is associated with a text description. The cornerstone of representation learning on TAGs lies in the seamless integration of textual semantics within individual nodes and the topological connections across nodes. Recent advancements in pre-trained language models (PLMs) and graph neural networks (GNNs) have facilitated effective learning on TAGs, garnering increased research interest. However, the absence of meaningful benchmark datasets and standardized evaluation procedures for TAGs has impeded progress in this field. In this paper, we propose CS-TAG, a comprehensive and diverse collection of challenging benchmark datasets for TAGs. The CS-TAG datasets are notably large in scale and encompass a wide range of domains, spanning from citation networks to purchase graphs. In addition to building the datasets, we conduct extensive benchmark experiments over CS-TAG with various learning paradigms, including PLMs, GNNs, PLM-GNN co-training methods, and the proposed novel topological pre-training of language models. In a nutshell, we provide an overview of the CS-TAG datasets, standardized evaluation procedures, and present baseline experiments. The entire CS-TAG project is publicly accessible at \url{https://github.com/sktsherlock/TAG-Benchmark}.

Sharing individual-level pandemic data is essential for accelerating the understanding of a disease. For example, COVID-19 data have been widely collected to support public health surveillance and research. In the United States, these data need to be de-identified before being released to the public due to privacy concerns. However, current data publishing approaches for individual-level pandemic data, such as those adopted by the U.S. Centers for Disease Control and Prevention (CDC), have not flexed over time to account for the dynamic nature of infection rates. Thus, the policies generated by these strategies may either raise privacy risks or impair the data utility (or usability). To optimize the tradeoff between privacy risk and data utility, we introduce a game theoretic model that adaptively generates policies to publish individual-level COVID-19 data according to infection dynamics. We model the data publishing process as a two-player Stackelberg game between a data publisher and a data recipient and then search for the best strategy for the publisher. In this game, we consider 1) the average accuracy of predicting future case counts for all demographic groups, and 2) the mutual information between the original data and the released data. We use COVID-19 case data from Vanderbilt University Medical Center from March 2020 to December 2021 to demonstrate our model and evaluate its effectiveness. The experimental results show that our game theoretic model outperforms all baseline approaches, including those adopted by CDC, while maintaining low privacy risk.

OBJECTIVE: Supporting public health research and the public's situational awareness during a pandemic requires continuous dissemination of infectious disease surveillance data. Legislation, such as the Health Insurance Portability and Accountability Act of 1996 and recent state-level regulations, permits sharing deidentified person-level data; however, current deidentification approaches are limited. Namely, they are inefficient, relying on retrospective disclosure risk assessments, and do not flex with changes in infection rates or population demographics over time. In this paper, we introduce a framework to dynamically adapt deidentification for near-real time sharing of person-level surveillance data. MATERIALS AND METHODS: The framework leverages a simulation mechanism, capable of application at any geographic level, to forecast the reidentification risk of sharing the data under a wide range of generalization policies. The estimates inform weekly, prospective policy selection to maintain the proportion of records corresponding to a group size less than 11 (PK11) at or below 0.1. Fixing the policy at the start of each week facilitates timely dataset updates and supports sharing granular date information. We use August 2020 through October 2021 case data from Johns Hopkins University and the Centers for Disease Control and Prevention to demonstrate the framework's effectiveness in maintaining the PK11 threshold of 0.01. RESULTS: When sharing COVID-19 county-level case data across all US counties, the framework's approach meets the threshold for 96.2% of daily data releases, while a policy based on current deidentification techniques meets the threshold for 32.3%. CONCLUSION: Periodically adapting the data publication policies preserves privacy while enhancing public health utility through timely updates and sharing epidemiologically critical features.

OBJECTIVE: Synthetic data are increasingly relied upon to share electronic health record (EHR) data while maintaining patient privacy. Current simulation methods can generate longitudinal data, but the results are unreliable for several reasons. First, the synthetic data drifts from the real data distribution over time. Second, the typical approach to quality assessment, which is based on the extent to which real records can be distinguished from synthetic records using a critic model, often fails to recognize poor simulation results. In this article, we introduce a longitudinal simulation framework, called LS-EHR, which addresses these issues. MATERIALS AND METHODS: LS-EHR enhances simulation through conditional fuzzing and regularization, rejection sampling, and prior knowledge embedding. We compare LS-EHR to the state-of-the-art using data from 60 000 EHRs from Vanderbilt University Medical Center (VUMC) and the All of Us Research Program. We assess discrimination between real and synthetic data over time. We evaluate the generation process and critic model using the area under the receiver operating characteristic curve (AUROC). For the critic, a higher value indicates a more robust model for quality assessment. For the generation process, a lower value indicates better synthetic data quality. RESULTS: The LS-EHR critic improves discrimination AUROC from 0.655 to 0.909 and 0.692 to 0.918 for VUMC and All of Us data, respectively. By using the new critic, the LS-EHR generation model reduces the AUROC from 0.909 to 0.758 and 0.918 to 0.806. CONCLUSION: LS-EHR can substantially improve the usability of simulated longitudinal EHR data.

OBJECTIVE: Deep learning models for clinical event forecasting (CEF) based on a patient's medical history have improved significantly over the past decade. However, their transition into practice has been limited, particularly for diseases with very low prevalence. In this paper, we introduce CEF-CL, a novel method based on contrastive learning to forecast in the face of a limited number of positive training instances. MATERIALS AND METHODS: CEF-CL consists of two primary components: (1) unsupervised contrastive learning for patient representation and (2) supervised transfer learning over the derived representation. We evaluate the new method along with state-of-the-art model architectures trained in a supervised manner with electronic health records data from Vanderbilt University Medical Center and the All of Us Research Program, covering 48 000 and 16 000 patients, respectively. We assess forecasting for over 100 diagnosis codes with respect to their area under the receiver operator characteristic curve (AUROC) and area under the precision-recall curve (AUPRC). We investigate the correlation between forecasting performance improvement and code prevalence via a Wald Test. RESULTS: CEF-CL achieved an average AUROC and AUPRC performance improvement over the state-of-the-art of 8.0%-9.3% and 11.7%-32.0%, respectively. The improvement in AUROC was negatively correlated with the number of positive training instances (P < .001). CONCLUSION: This investigation indicates that clinical event forecasting can be improved significantly through contrastive representation learning, especially when the number of positive training instances is small.

OBJECTIVE: In intensive care units (ICUs), a patient's brain function status can shift from a state of acute brain dysfunction (ABD) to one that is ABD-free and vice versa, which is challenging to forecast and, in turn, hampers the allocation of hospital resources. We aim to develop a machine learning model to predict next-day brain function status changes. MATERIALS AND METHODS: Using multicenter prospective adult cohorts involving medical and surgical ICU patients from 2 civilian and 3 Veteran Affairs hospitals, we trained and externally validated a light gradient boosting machine to predict brain function status changes. We compared the performances of the boosting model against state-of-the-art models-an ABD predictive model and its variants. We applied Shapley additive explanations to identify influential factors to develop a compact model. RESULTS: There were 1026 critically ill patients without evidence of prior major dementia, or structural brain diseases, from whom 12 295 daily transitions (ABD: 5847 days; ABD-free: 6448 days) were observed. The boosting model achieved an area under the receiver-operating characteristic curve (AUROC) of 0.824 (95% confidence interval [CI], 0.821-0.827), compared with the state-of-the-art models of 0.697 (95% CI, 0.693-0.701) with P < .001. Using 13 identified top influential factors, the compact model achieved 99.4% of the boosting model on AUROC. The boosting and the compact models demonstrated high generalizability in external validation by achieving an AUROC of 0.812 (95% CI, 0.812-0.813). CONCLUSION: The inputs of the compact model are based on several simple questions that clinicians can quickly answer in practice, which demonstrates the model has direct prospective deployment potential into clinical practice, aiding in critical hospital resource allocation.

OBJECTIVE: Simulating electronic health record data offers an opportunity to resolve the tension between data sharing and patient privacy. Recent techniques based on generative adversarial networks have shown promise but neglect the temporal aspect of healthcare. We introduce a generative framework for simulating the trajectory of patients' diagnoses and measures to evaluate utility and privacy. MATERIALS AND METHODS: The framework simulates date-stamped diagnosis sequences based on a 2-stage process that 1) sequentially extracts temporal patterns from clinical visits and 2) generates synthetic data conditioned on the learned patterns. We designed 3 utility measures to characterize the extent to which the framework maintains feature correlations and temporal patterns in clinical events. We evaluated the framework with billing codes, represented as phenome-wide association study codes (phecodes), from over 500 000 Vanderbilt University Medical Center electronic health records. We further assessed the privacy risks based on membership inference and attribute disclosure attacks. RESULTS: The simulated temporal sequences exhibited similar characteristics to real sequences on the utility measures. Notably, diagnosis prediction models based on real versus synthetic temporal data exhibited an average relative difference in area under the ROC curve of 1.6% with standard deviation of 3.8% for 1276 phecodes. Additionally, the relative difference in the mean occurrence age and time between visits were 4.9% and 4.2%, respectively. The privacy risks in synthetic data, with respect to the membership and attribute inference were negligible. CONCLUSION: This investigation indicates that temporal diagnosis code sequences can be simulated in a manner that provides utility and respects privacy.

OBJECTIVE: Hospital capacity management depends on accurate real-time estimates of hospital-wide discharges. Estimation by a clinician requires an excessively large amount of effort and, even when attempted, accuracy in forecasting next-day patient-level discharge is poor. This study aims to support next-day discharge predictions with machine learning by incorporating electronic health record (EHR) audit log data, a resource that captures EHR users' granular interactions with patients' records by communicating various semantics and has been neglected in outcome predictions. MATERIALS AND METHODS: This study focused on the EHR data for all adults admitted to Vanderbilt University Medical Center in 2019. We learned multiple advanced models to assess the value that EHR audit log data adds to the daily prediction of discharge likelihood within 24 h and to compare different representation strategies. We applied Shapley additive explanations to identify the most influential types of user-EHR interactions for discharge prediction. RESULTS: The data include 26 283 inpatient stays, 133 398 patient-day observations, and 819 types of user-EHR interactions. The model using the count of each type of interaction in the recent 24 h and other commonly used features, including demographics and admission diagnoses, achieved the highest area under the receiver operating characteristics (AUROC) curve of 0.921 (95% CI: 0.919-0.923). By contrast, the model lacking user-EHR interactions achieved a worse AUROC of 0.862 (0.860-0.865). In addition, 10 of the 20 (50%) most influential factors were user-EHR interaction features. CONCLUSION: EHR audit log data contain rich information such that it can improve hospital-wide discharge predictions.

Detection of malicious behavior is a fundamental problem in security. One of the major challenges in using detection systems in practice is in dealing with an overwhelming number of alerts that are triggered by normal behavior (the so-called false positives), obscuring alerts resulting from actual malicious activities. We introduce a novel approach for computing a policy for prioritizing alerts using adversarial reinforcement learning. Our approach assumes that the attacker knows the full state of the detection system and the defender's alert prioritization policy, and will dynamically choose an optimal attack. The first step of our approach is to capture the interaction between the defender and attacker in a game theoretic model. To tackle the computational complexity of solving this game to obtain a dynamic stochastic alert prioritization policy, we propose an adversarial reinforcement learning framework. In this framework, we use neural reinforcement learning to compute best response policies for both the defender and the adversary to an arbitrary stochastic policy of the other. We then use these in a double-oracle framework to obtain an approximate equilibrium of the game, which in turn yields a robust stochastic policy for the defender. We use case studies in network intrusion and fraud detection to demonstrate that our approach is effective in creating robust alert prioritization policies.1

Routine operational use of sensitive data is often governed by law and regulation. For instance, in the medical domain, there are various statues at the state and federal level that dictate who is permitted to work with patients' records and under what conditions. To screen for potential privacy breaches, logging systems are usually deployed to trigger alerts whenever a suspicious access is detected. However, such mechanisms are often inefficient because 1) the vast majority of triggered alerts are false positives, 2) small budgets make it unlikely that a real attack will be detected, and 3) attackers can behave strategically, such that traditional auditing mechanisms cannot easily catch them. To improve efficiency, information systems may invoke signaling, so that whenever a suspicious access request occurs, the system can, in real time, warn the user that the access may be audited. Then, at the close of a finite period, a selected subset of suspicious accesses are audited. This gives rise to an online problem in which one needs to determine 1) whether a warning should be triggered and 2) the likelihood that the data request event will be audited. In this paper, we formalize this auditing problem as a Signaling Audit Game (SAG), in which we model the interactions between an auditor and an attacker in the context of signaling and the usability cost is represented as a factor of the auditor's payoff. We study the properties of its Stackelberg equilibria and develop a scalable approach to compute its solution. We show that a strategic presentation of warnings adds value in that SAGs realize significantly higher utility for the auditor than systems without signaling. We perform a series of experiments with 10 million real access events, containing over 26K alerts, from a large academic medical center to illustrate the value of the proposed auditing model and the consistency of its advantages over existing baseline methods.

OBJECTIVE: Electronic medical records (EMRs) can support medical research and discovery, but privacy risks limit the sharing of such data on a wide scale. Various approaches have been developed to mitigate risk, including record simulation via generative adversarial networks (GANs). While showing promise in certain application domains, GANs lack a principled approach for EMR data that induces subpar simulation. In this article, we improve EMR simulation through a novel pipeline that (1) enhances the learning model, (2) incorporates evaluation criteria for data utility that informs learning, and (3) refines the training process. MATERIALS AND METHODS: We propose a new electronic health record generator using a GAN with a Wasserstein divergence and layer normalization techniques. We designed 2 utility measures to characterize similarity in the structural properties of real and simulated EMRs in the original and latent space, respectively. We applied a filtering strategy to enhance GAN training for low-prevalence clinical concepts. We evaluated the new and existing GANs with utility and privacy measures (membership and disclosure attacks) using billing codes from over 1 million EMRs at Vanderbilt University Medical Center. RESULTS: The proposed model outperformed the state-of-the-art approaches with significant improvement in retaining the nature of real records, including prediction performance and structural properties, without sacrificing privacy. Additionally, the filtering strategy achieved higher utility when the EMR training dataset was small. CONCLUSIONS: These findings illustrate that EMR simulation through GANs can be substantially improved through more appropriate training, modeling, and evaluation criteria.

The quantity of personal data that is collected, stored, and subsequently processed continues to grow rapidly. Given its sensitivity, ensuring privacy protections has become a necessary component of database management. To enhance protection, a number of mechanisms have been developed, such as audit logging and alert triggers, which notify administrators about suspicious activities. However, this approach is limited. First, the volume of alerts is often substantially greater than the auditing capabilities of organizations. Second, strategic attackers can attempt to disguise their actions or carefully choose targets, thus hide illicit activities. In this article, we introduce an auditing approach that accounts for adversarial behavior by (1) prioritizing the order in which types of alerts are investigated and (2) providing an upper bound on how much resource to allocate for each type. Specifically, we model the interaction between a database auditor and attackers as a Stackelberg game. We show that even a highly constrained version of such problem is NP-Hard. Then, we introduce a method that combines linear programming, column generation, and heuristic searching to derive an auditing policy. On the synthetic data, we perform an extensive evaluation on the approximation degree of our solution with the optimal one. The two real datasets, (1) 1.5 months of audit logs from Vanderbilt University Medical Center and (2) a publicly available credit card application dataset, are used to test the policy-searching performance. The findings demonstrate the effectiveness of the proposed methods for searching the audit strategies, and our general approach significantly outperforms non-game-theoretic baselines.

A wide variety of mechanisms, such as alert triggers and auditing routines, have been developed to notify administrators about types of suspicious activities in the daily use of large databases of personal and sensitive information. However, such mechanisms are limited in that: 1) the volume of such alerts is often substantially greater than the auditing capabilities of budget-constrained organizations and 2) strategic attackers may disguise their actions or carefully choose which records they touch, thus evading auditing routines. To address these problems, we introduce a novel approach to database auditing that explicitly accounts for adversarial behavior by 1) prioritizing the order in which types of alerts are investigated and 2) providing an upper bound on how much budget to allocate for auditing each alert type. We model the interaction between a database auditor and potential attackers as a Stackelberg game in which the auditor chooses an auditing policy and attackers choose which records in a database to target. We further introduce an efficient approach that combines linear programming, column generation, and heuristic search to derive an auditing policy, in the form of a mixed strategy. We assess the performance of the policy selection method using a publicly available credit card application dataset, the results of which indicate that our method produces high-quality database audit policies, significantly outperforming baselines that are not based in a game theoretic framing.