A systems approach to designing effective clinical trials using simulations

Approximating Functions with Approximate Privacy for Applications in Signal Estimation and Learning

Large corporations, government entities and institutions such as hospitals and census bureaus routinely collect our personal and sensitive information for providing services. A key technological challenge is designing algorithms for these services that provide useful results, while simultaneously maintaining the privacy of the individuals whose data are being shared. Differential privacy (DP) is a cryptographically motivated and mathematically rigorous approach for addressing this challenge. Under DP, a randomized algorithm provides privacy guarantees by approximating the desired functionality, leading to a privacy-utility trade-off. Strong (pure DP) privacy guarantees are often costly in terms of utility. Motivated by the need for a more efficient mechanism with better privacy-utility trade-off, we propose Gaussian FM, an improvement to the functional mechanism (FM) that offers higher utility at the expense of a weakened (approximate) DP guarantee. We analytically show that the proposed Gaussian FM algorithm can offer orders of magnitude smaller noise compared to the existing FM algorithms. We further extend our Gaussian FM algorithm to decentralized-data settings by incorporating the CAPE protocol and propose capeFM. Our method can offer the same level of utility as its centralized counterparts for a range of parameter choices. We empirically show that our proposed algorithms outperform existing state-of-the-art approaches on synthetic and real datasets.

Optimizing warfarin dosing using deep reinforcement learning

Warfarin is a widely used anticoagulant, and has a narrow therapeutic range. Dosing of warfarin should be individualized, since slight overdosing or underdosing can have catastrophic or even fatal consequences. Despite much research on warfarin dosing, current dosing protocols do not live up to expectations, especially for patients sensitive to warfarin. We propose a deep reinforcement learning-based dosing model for warfarin. To overcome the issue of relatively small sample sizes in dosing trials, we use a Pharmacokinetic/ Pharmacodynamic (PK/PD) model of warfarin to simulate dose-responses of virtual patients. Applying the proposed algorithm on virtual test patients shows that this model outperforms a set of clinically accepted dosing protocols by a wide margin. We tested the robustness of our dosing protocol on a second PK/PD model and showed that its performance is comparable to the set of baseline protocols.

Optimizing the dynamic treatment regime of in-hospital warfarin anticoagulation in patients after surgical valve replacement using reinforcement learning

OBJECTIVE

Warfarin anticoagulation management requires sequential decision-making to adjust dosages based on patients' evolving states continuously. We aimed to leverage reinforcement learning (RL) to optimize the dynamic in-hospital warfarin dosing in patients after surgical valve replacement (SVR).

MATERIALS AND METHODS

10 408 SVR cases with warfarin dosage-response data were retrospectively collected to develop and test an RL algorithm that can continuously recommend daily warfarin doses based on patients' evolving multidimensional states. The RL algorithm was compared with clinicians' actual practice and other machine learning and clinical decision rule-based algorithms. The primary outcome was the ratio of patients without in-hospital INRs >3.0 and the INR at discharge within the target range (1.8-2.5) (excellent responders). The secondary outcomes were the safety responder ratio (no INRs >3.0) and the target responder ratio (the discharge INR within 1.8-2.5).

RESULTS

In the test set (n = 1260), the excellent responder ratio under clinicians' guidance was significantly lower than the RL algorithm: 41.6% versus 80.8% (relative risk [RR], 0.51; 95% confidence interval [CI], 0.48-0.55), also the safety responder ratio: 83.1% versus 99.5% (RR, 0.83; 95% CI, 0.81-0.86), and the target responder ratio: 49.7% versus 81.1% (RR, 0.61; 95% CI, 0.58-0.65). The RL algorithms performed significantly better than all the other algorithms. Compared with clinicians' actual practice, the RL-optimized INR trajectory reached and maintained within the target range significantly faster and longer.

DISCUSSION

RL could offer interactive, practical clinical decision support for sequential decision-making tasks and is potentially adaptable for varied clinical scenarios. Prospective validation is needed.

CONCLUSION

An RL algorithm significantly optimized the post-operation warfarin anticoagulation quality compared with clinicians' actual practice, suggesting its potential for challenging sequential decision-making tasks.

Study on the therapeutic material basis and effect of Acanthopanax senticosus (Rupr. et Maxim.) Harms leaves in the treatment of ischemic stroke by PK-PD analysis based on online microdialysis-LC-MS/MS method

Leaves of Acanthopanax senticosus (Rupr. et Maxim.) Harms (ASL) have revealed significant biological activity in the treatment of ischemic stroke diseases. However, there was no in-depth study of the therapeutic material basis and effect of ASL from the pharmacokinetics-pharmacodynamics (PK-PD) analysis level. In this study, a method based on microdialysis coupled with ultra-performance liquid chromatography combined with triple quadruple mass spectrometry (MD-UPLC-QQQ-MS) was established to simultaneously and continuously collect and quantify the active compounds and endogenous neuroactive substances related to therapeutic effect in plasma and hippocampus of fully awake ischemic stroke rats. The acquired data were analyzed by the PK-PD analysis method. It was found that hyperoside, quercitrin, quercetin, and caffeic acid could pass through the blood-brain barrier, and quercetin needed a longer intake time than quercitrin and hyperoside, but the passage rate was higher. The exposure of the four compounds in the hippocampus affected the contents of seven neuroactive substances in different ways and was depicted graphically (concentration-time effect). In addition, the study found that the brain index and brain water content of ischemic stroke rats were significantly reduced after the oral administration of ASL. ASL observably regulated the content or activity of six important biochemical indexes in rats. On the one hand, this study verified that ASL could regulate ischemic stroke in many aspects. On the other hand, a visualized method to express the relationship between pharmacokinetics and pharmacodynamics in the hippocampus of cerebral ischemic areas was established. This research gives a hand to the study on the therapeutic material basis and effect of traditional Chinese medicine mechanism.

Personalized Anticoagulation: Optimizing Warfarin Management Using Genetics and Simulated Clinical Trials

BACKGROUND

Clinical trials testing pharmacogenomic-guided warfarin dosing for patients with atrial fibrillation have demonstrated conflicting results. Non-vitamin K antagonist oral anticoagulants are expensive and contraindicated for several conditions. A strategy optimizing anticoagulant selection remains an unmet clinical need.

METHODS AND RESULTS

Characteristics from 14 206 patients with atrial fibrillation were integrated into a validated warfarin clinical trial simulation framework using iterative Bayesian network modeling and a pharmacokinetic-pharmacodynamic model. Individual dose-response for patients was simulated for 5 warfarin protocols-a fixed-dose protocol, a clinically guided protocol, and 3 increasingly complex pharmacogenomic-guided protocols. For each protocol, a complexity score was calculated using the variables predicting warfarin dose and the number of predefined international normalized ratio (INR) thresholds for each adjusted dose. Study outcomes included optimal time in therapeutic range ≥65% and clinical events. A combination of age and genotype identified different optimal protocols for various subpopulations. A fixed-dose protocol provided well-controlled INR only in normal responders ≥65, whereas for normal responders <65 years old, a clinically guided protocol was necessary to achieve well-controlled INR. Sensitive responders ≥65 and <65 and highly sensitive responders ≥65 years old required pharmacogenomic-guided protocols to achieve well-controlled INR. However, highly sensitive responders <65 years old did not achieve well-controlled INR and had higher associated clinical events rates than other subpopulations.

CONCLUSIONS

Under the assumptions of this simulation, patients with atrial fibrillation can be triaged to an optimal warfarin therapy protocol by age and genotype. Clinicians should consider alternative anticoagulation therapy for patients with suboptimal outcomes under any warfarin protocol.

Optimal decision support rules improve personalize warfarin treatment outcomes

We tested optimization-based approaches to generate decision support rules used to improve personalized warfarin treatment based on clinical and genetic characteristics. Our approach simulated warfarin treatment outcomes using five existing treatment plans for clinical avatars (virtual patients). We used individual clinical avatar Time-in-Therapeutic-Range to represent the two-sided adverse risk to bleeding (over dosed - above therapeutic range) and thrombosis (under dosed - below therapeutic range) and as the objective function in the optimization to minimize overall risk. A series of optimization approaches demonstrate that correctly selected decision rules matched to particularly characterized patients produce treatment plans that minimize risk. Finally, a decision tree algorithm was used to produce decision rules, each of which indicated a specific treatment plan that optimally reduce risks for a patient subgroup. The optimization approach minimizes entropy/impurity property thus producing rules that identify treatment plans that minimize overall adverse risks for the largest possible patient subgroups.

Privacy in Pharmacogenetics: An End-to-End Case Study of Personalized Warfarin Dosing

We initiate the study of privacy in pharmacogenetics, wherein machine learning models are used to guide medical treatments based on a patient's genotype and background. Performing an in-depth case study on privacy in personalized warfarin dosing, we show that suggested models carry privacy risks, in particular because attackers can perform what we call model inversion: an attacker, given the model and some demographic information about a patient, can predict the patient's genetic markers. As differential privacy (DP) is an oft-proposed solution for medical settings such as this, we evaluate its effectiveness for building private versions of pharmacogenetic models. We show that DP mechanisms prevent our model inversion attacks when the privacy budget is carefully selected. We go on to analyze the impact on utility by performing simulated clinical trials with DP dosing models. We find that for privacy budgets effective at preventing attacks, patients would be exposed to increased risk of stroke, bleeding events, and mortality. We conclude that current DP mechanisms do not simultaneously improve genomic privacy while retaining desirable clinical efficacy, highlighting the need for new mechanisms that should be evaluated in situ using the general methodology introduced by our work.

Next generation phenotyping using the unified medical language system

Background

Structured information within patient medical records represents a largely untapped treasure trove of research data. In the United States, privacy issues notwithstanding, this has recently become more accessible thanks to the increasing adoption of electronic health records (EHR) and health care data standards fueled by the Meaningful Use legislation. The other side of the coin is that it is now becoming increasingly more difficult to navigate the profusion of many disparate clinical terminology standards, which often span millions of concepts.

Objective

The objective of our study was to develop a methodology for integrating large amounts of structured clinical information that is both terminology agnostic and able to capture heterogeneous clinical phenotypes including problems, procedures, medications, and clinical results (such as laboratory tests and clinical observations). In this context, we define phenotyping as the extraction of all clinically relevant features contained in the EHR.

Methods

The scope of the project was framed by the Common Meaningful Use (MU) Dataset terminology standards; the Systematized Nomenclature of Medicine Clinical Terms (SNOMED CT), RxNorm, the Logical Observation Identifiers Names and Codes (LOINC), the Current Procedural Terminology (CPT), the Health care Common Procedure Coding System (HCPCS), the International Classification of Diseases Ninth Revision Clinical Modification (ICD-9-CM), and the International Classification of Diseases Tenth Revision Clinical Modification (ICD-10-CM). The Unified Medical Language System (UMLS) was used as a mapping layer among the MU ontologies. An extract, load, and transform approach separated original annotations in the EHR from the mapping process and allowed for continuous updates as the terminologies were updated. Additionally, we integrated all terminologies into a single UMLS derived ontology and further optimized it to make the relatively large concept graph manageable.

Results

The initial evaluation was performed with simulated data from the Clinical Avatars project using 100,000 virtual patients undergoing a 90 day, genotype guided, warfarin dosing protocol. This dataset was annotated with standard MU terminologies, loaded, and transformed using the UMLS. We have deployed this methodology to scale in our in-house analytics platform using structured EHR data for 7931 patients (12 million clinical observations) treated at the Froedtert Hospital. A demonstration limited to Clinical Avatars data is available on the Internet using the credentials user “jmirdemo” and password “jmirdemo”.

Conclusions

Despite its inherent complexity, the UMLS can serve as an effective interface terminology for many of the clinical data standards currently used in the health care domain.

How often should we monitor for reliable detection of atrial fibrillation recurrence? Efficiency considerations and implications for study design

Objective

Although atrial fibrillation (AF) recurrence is unpredictable in terms of onset and duration, current intermittent rhythm monitoring (IRM) diagnostic modalities are short-termed and discontinuous. The aim of the present study was to investigate the necessary IRM frequency required to reliably detect recurrence of various AF recurrence patterns.

Methods

The rhythm histories of 647 patients (mean AF burden: 12±22% of monitored time; 687 patient-years) with implantable continuous monitoring devices were reconstructed and analyzed. With the use of computationally intensive simulation, we evaluated the necessary IRM frequency to reliably detect AF recurrence of various AF phenotypes using IRM of various durations.

Results

The IRM frequency required for reliable AF detection depends on the amount and temporal aggregation of the AF recurrence (p<0.0001) as well as the duration of the IRM (p<0.001). Reliable detection (>95% sensitivity) of AF recurrence required higher IRM frequencies (>12 24-hour; >6 7-day; >4 14-day; >3 30-day IRM per year; p<0.0001) than currently recommended. Lower IRM frequencies will under-detect AF recurrence and introduce significant bias in the evaluation of therapeutic interventions. More frequent but of shorter duration, IRMs (24-hour) are significantly more time effective (sensitivity per monitored time) than a fewer number of longer IRM durations (p<0.0001).

Conclusions

Reliable AF recurrence detection requires higher IRM frequencies than currently recommended. Current IRM frequency recommendations will fail to diagnose a significant proportion of patients. Shorter duration but more frequent IRM strategies are significantly more efficient than longer IRM durations.

Clinical Trial Registration URL

Unique identifier: NCT00806689.

Simulating sequential multiple assignment randomized trials to generate optimal personalized warfarin dosing strategies

Background

Due to the cost and complexity of conducting a sequential multiple assignment randomized trial (SMART), it is desirable to pre-define a small number of personalized regimes to study.

Purpose

We proposed a simulation-based approach to studying personalized dosing strategies in contexts for which a therapeutic agent’s pharmacokinetic and pharmacodynamics properties are well understood. We take dosing of warfarin as a case study, as its properties are well understood. We consider a SMART in which there are five intervention points in which dosing may be modified, following a loading phase of treatment.

Methods

Realistic SMARTs are simulated, and two methods of analysis, G-estimation and Q-learning, are used to assess potential personalized dosing strategies.

Results

In settings where outcome modelling may be complex due to the highly non-linear nature of the pharmacokinetic and pharmacodynamics mechanisms of the therapeutic agent, G-estimation provides for which the more promising method of estimating an optimal dosing strategy. Used in combination with the simulated SMARTs, we were able to improve simulated patient outcomes and suggest which patient characteristics were needed to best individually tailor dosing. In particular, our simulations suggest that current dosing should be determined by an individual’s current coagulation time as measured by the international normalized ratio (INR), their last measured INR, and their last dose. Tailoring treatment only based on current INR and last warfarin dose provided inferior control of INR over the course of the trial.

Limitations

The ability of the simulated SMARTs to suggest optimal personalized dosing strategies relies on the pharmacokinetic and pharmacodynamic models used to generate the hypothetical patient profiles. This approach is best suited to therapeutic agents whose effects are well studied.

Conclusion

Prior to investing in a complex randomized trial that involves sequential treatment allocations, simulations should be used where possible in order to guide which dosing strategies to evaluate.