The Unseen Hand: AI-Based Prescribing Decision Support Tools and the Evaluation of Drug Safety and Effectiveness.

The use of artificial intelligence (AI)-based tools to guide prescribing decisions is full of promise and may enhance patient outcomes. These tools can perform actions such as choosing the 'safest' medication, choosing between competing medications, promoting de-prescribing or even predicting non-adherence. These tools can exist in a variety of formats; for example, they may be directly integrated into electronic medical records or they may exist in a stand-alone website accessible by a web browser. One potential impact of these tools is that they could manipulate our understanding of the benefit-risk of medicines in the real world. Currently, the benefit risk of approved medications is assessed according to carefully planned agreements covering spontaneous reporting systems and planned surveillance studies. But AI-based tools may limit or even block prescription to high-risk patients or prevent off-label use. The uptake and temporal availability of these tools may be uneven across healthcare systems and geographies, creating artefacts in data that are difficult to account for. It is also hard to estimate the 'true impact' that a tool had on a prescribing decision. International borders may also be highly porous to these tools, especially in cases where tools are available over the web. These tools already exist, and their use is likely to increase in the coming years. How they can be accounted for in benefit-risk decisions is yet to be seen.

A new approach to identifying patients with elevated risk for Fabry disease using a machine learning algorithm.

BACKGROUND: Fabry disease (FD) is a rare genetic disorder characterized by glycosphingolipid accumulation and progressive damage across multiple organ systems. Due to its heterogeneous presentation, the condition is likely significantly underdiagnosed. Several approaches, including provider education efforts and newborn screening, have attempted to address underdiagnosis of FD across the age spectrum, with limited success. Artificial intelligence (AI) methods present another option for improving diagnosis. These methods isolate common health history patterns among patients using longitudinal real-world data, and can be particularly useful when patients experience nonspecific, heterogeneous symptoms over time. In this study, the performance of an AI tool in identifying patients with FD was analyzed. The tool was calibrated using de-identified health record data from a large cohort of nearly 5000 FD patients, and extracted phenotypic patterns from these records. The tool then used this FD pattern information to make individual-level estimates of FD in a testing dataset. Patterns were reviewed and confirmed with medical experts. RESULTS: The AI tool demonstrated strong analytic performance in identifying FD patients. In out-of-sample testing, it achieved an area under the receiver operating characteristic curve (AUROC) of 0.82. Strong performance was maintained when testing on male-only and female-only cohorts, with AUROCs of 0.83 and 0.82 respectively. The tool identified small segments of the population with greatly increased prevalence of FD: in the 1% of the population identified by the tool as at highest risk, FD was 23.9 times more prevalent than in the population overall. The AI algorithm used hundreds of phenotypic signals to make predictions and included both familiar symptoms associated with FD (e.g. renal manifestations) as well as less well-studied characteristics. CONCLUSIONS: The AI tool analyzed in this study performed very well in identifying Fabry disease patients using structured medical history data. Performance was maintained in all-male and all-female cohorts, and the phenotypic manifestations of FD highlighted by the tool were reviewed and confirmed by clinical experts in the condition. The platform's analytic performance, transparency, and ability to generate predictions based on existing real-world health data may allow it to contribute to reducing persistent underdiagnosis of Fabry disease.

A Bayesian Belief Network Model Assessing the Risk to Wastewater Workers of Contracting Ebola Virus Disease During an Outbreak.

During an outbreak of Ebola virus disease (EVD), hospitals' connections to municipal wastewater systems may provide a path for patient waste bearing infectious viral particles to pass from the hospital into the wastewater treatment system, potentially posing risks to sewer and wastewater workers. To quantify these risks, we developed a Bayesian belief network model incorporating data on virus behavior and survival along with structural characteristics of hospitals and wastewater treatment systems. We applied the model to assess risks under several different scenarios of workers' exposure to wastewater for a wastewater system typical of a mid-sized U.S. city. The model calculates a median daily risk of developing EVD of approximately 6.1×10-12 (90% confidence interval: 1.0×10-12 to 5.4×10-9 ; mean 1.8×10-6 ) when no prior exposure conditions are specified. Under a worst-case scenario in which a worker stationed in the sewer adjacent to the hospital accidentally ingests several drops (0.35 mL) of wastewater, median risk is 5.8×10-4 (90% CI: 8.8×10-7 to 9.5×10-2 ; mean 3.2×10-2 ) . Disinfection of patient waste with peracetic acid for 15 minutes prior to flushing decreases the estimated median risk to 3.8×10-7 (90% CI: 4.1×10-9 to 8.6×10-5 ; mean 2.9×10-5 ). The results suggest that requiring hospitals to disinfect EVD patient waste prior to flushing may be advisable. The modeling framework can provide insight into managing patient waste during future outbreaks of highly virulent infectious pathogens.

Advancing Dose-Response Assessment Methods for Environmental Regulatory Impact Analysis: A Bayesian Belief Network Approach Applied to Inorganic Arsenic.

Dose-response functions used in regulatory risk assessment are based on studies of whole organisms and fail to incorporate genetic and metabolomic data. Bayesian belief networks (BBNs) could provide a powerful framework for incorporating such data, but no prior research has examined this possibility. To address this gap, we develop a BBN-based model predicting birthweight at gestational age from arsenic exposure via drinking water and maternal metabolic indicators using a cohort of 200 pregnant women from an arsenic-endemic region of Mexico. We compare BBN predictions to those of prevailing slope-factor and reference-dose approaches. The BBN outperforms prevailing approaches in balancing false-positive and false-negative rates. Whereas the slope-factor approach had 2% sensitivity and 99% specificity and the reference-dose approach had 100% sensitivity and 0% specificity, the BBN's sensitivity and specificity were 71% and 30%, respectively. BBNs offer a promising opportunity to advance health risk assessment by incorporating modern genetic and metabolomic data.