Bayesian Machine Learning Model Guiding Iterative, Personalized Anticoagulant Dosing Decision-Making

Abstract

Background

Anticoagulation in atrial fibrillation often relies on a fixed dose and infrequent dose adjustment and does not incorporate patient preferences for risks of stroke, bleeding, and death.

Objectives

We developed a Bayesian machine learning model (Adele) to guide individualized long-term dosing accounting for patient preferences.

Methods

Adele is a Bayesian competing-risk, multistate hazard model trained on 5,380 edoxaban-treated patients with atrial fibrillation with pharmacokinetic (PK) data from the ENGAGE AF-TIMI 48 (Effective Anticoagulation with Factor Xa Next Generation in Atrial Fibrillation–TIMI 48) trial. Patient PK and baseline data were used to estimate personalized continual risk prediction. The concordance index compared predictive accuracy of Adele with standard Kaplan-Meier estimators. Positive concordance index values (%) indicate better predictive accuracy for Adele. In 2 patient examples, we demonstrate Adele's ability to predict 3-year event frequencies at randomization and following intercurrent events identifying optimal dosing across three edoxaban doses (60mg, 30mg, 15mg) based on hypothetical outcome preferences.

Results

Adele outperformed Kaplan-Meier estimators, improving 3-year prediction for cardiovascular death by +12.1%, disability by +11.8%, major gastrointestinal bleeding by +13.7%, and ischemic stroke (IS) by +3.3%. Adele demonstrated dynamic risk prediction capabilities, with future event probabilities shifting following intracranial hemorrhage or IS. In patient A (80-yr female; 52 kg), preference-weighted event rates were lowest at 15 mg and 30 mg, irrespective of the preference to avoid cardiovascular death, major bleeding, or disability. In patient B (72-yr male; 80 kg), preference-weighted event rates were lowest at 60 mg, irrespective of the preference to avoid death or IS/intracranial hemorrhage, and instead lowest at 15 mg for disability avoidance.

Conclusions

Adele represents a state-of-the-art Bayesian framework, using clinical factors and PK data, to enable longitudinally adaptive, patient-centric, preference-weighted predictions.

Central Illustration

Modern clinical practice relies on evidence from population-based trials, requiring physicians to apply these findings to treat individual patients.^1,2 However, population-level results often fail to capture critical nuances at the patient level: interaction of multiple complex risk factors, changes in patient characteristics including intercurrent events during follow-up, and individual patient benefit–risk preferences.1, 2, 3, 4 Furthermore, regulatory-approved dosing recommendations reflect efficacy-safety trade-offs optimized for study population averages rather than individual patients.¹

The future of medicine demands tools that address these limitations by: 1) providing dynamic, personalized risk estimates incorporating multiple risk factors and drug exposure; 2) enabling dynamic prediction of subsequent clinical events (eg, probability of cardiovascular [CV] fatality following a bleeding event); and 3) facilitating shared decision-making that incorporates patient preference-driven dose selection.3, 4, 5 Regulatory agencies and clinical practice guidelines have increasingly recognized the importance of precision medicine approaches, encouraging the development of tools that can support personalized dosing strategies.6, 7, 8

Atrial fibrillation (AF) affects around 59 million individuals worldwide, including approximately 10 million Americans, with the global prevalence expected to double by 2050 due to population aging.⁹ AF is typically a lifelong condition that significantly increases the risk of ischemic events, particularly thromboembolic stroke.¹⁰ Continuous changes in patient risk and preference require dynamic adaptation to optimize anticoagulation management.^5,11 Current AF guidelines rely on established risk stratification tools such as CHA₂DS₂-VA (defined as congestive heart failure, hypertension, age ≥75 years [doubled], diabetes mellitus, prior stroke or transient ischemic attack or thromboembolism [doubled], vascular disease) for ischemic stroke (IS) risk and HAS-BLED (defined as hypertension, abnormal renal or liver function, stroke, bleeding tendency or predisposition, labile international normalized ratios, age >65 years, and drugs or alcohol use) for bleeding risk assessment.^4,12 However, these tools provide static risk estimates at treatment initiation that do not account for changing patient circumstances, such as intercurrent events, or individual patient preferences regarding ischemic vs bleeding outcomes.

Direct oral anticoagulants (DOACs), including the once-daily oral factor Xa inhibitor edoxaban, represent the standard of care for reducing thromboembolic risk in patients with AF.^4,12 The phase 3 ENGAGE AF-TIMI-48 (Effective Anticoagulation with Factor Xa Next Generation in Atrial Fibrillation–TIMI 48) trial (NCT00781391) demonstrated edoxaban's noninferiority to warfarin for IS prevention, while significantly reducing CV death and major bleeding rates.¹³ In clinical practice, DOAC treatment decisions must carefully balance IS prevention against bleeding risk, requiring a comprehensive understanding of individual patient risk profiles—including disease history and comorbidities—to guide optimal dose selection.⁴ The ENGAGE AF-TIMI 48 trial data set provides an invaluable resource for developing such precision medicine approaches.¹³

Machine learning (ML), artificial intelligence, and state-of-the-art Bayesian approaches offer transformative potential for AF management by addressing limitations inherent to traditional risk stratification approaches. Bayesian ML models can integrate high-dimensional data—including pharmacokinetic (PK) exposures, time-varying patient characteristics, and sequential clinical events—to identify previously unrecognized patterns and generate personalized predictions that evolve as patients experience intercurrent events.^5,14,15 Bayesian approaches also offer particular advantages for clinical decision support: principled quantification of prediction uncertainty, natural incorporation of prior evidence from clinical trials, and the ability to update risk estimates dynamically as new patient data accumulate. Such approaches will enable dynamic treatment adaptation based on personalized risk predictions and patient preferences.^5,14,15 These capabilities are especially relevant for edoxaban dose optimization, where treatment decisions must balance competing risks (IS prevention vs bleeding) that vary substantially across patients and evolve throughout the treatment course.^4,5,9 Recent advances in survival modeling have included techniques such as random survival forests, gradient boosting, and deep learning-based survival models (eg, DeepSurv, DeepHit). These approaches aim to capture complex, nonlinear relationships and time-to-event dynamics, but at the present, often require large data sets and can lack clear interpretability—limiting their adoption in clinical practice.¹⁶

Adele (named in honor of Myra Adele Logan, the pioneering African American surgeon who was the first woman to perform successful open-heart surgery) is a Bayesian, joint competing-risk, multistate exposure-hazard model trained on ENGAGE AF-TIMI 48 trial data.17, 18, 19, 20, 21, 22 Rather than a singular artificial intelligence innovation, Adele represents an evolution within the precision dosing continuum—building upon foundational PK/PD modeling and advancing toward real-time adaptive clinical guidance. This state-of-the-art Bayesian framework enables longitudinally adaptive, patient-centric event predictions and incorporates individual patient preferences, supporting dynamic dose optimization as patient risk profiles and clinical circumstances evolve. The primary objective was to evaluate Adele's capability as a novel methodological tool to support shared decision-making for edoxaban dose optimization and iteratively and longitudinally assess individual risk throughout treatment; potentially reduce specific events; enhance understanding of patient transitions across edoxaban regimens; and incorporate individual outcome preferences to inform dosing decisions.

Methods

ADELE: A bayesian multistate model with competing risks

Adele is a Bayesian competing-risk, multistate hazard model, developed in 5,380 edoxaban-treated patients with complete PK data from the ENGAGE AF-TIMI 48 trial.¹³ The model was trained on 80% of the population (n = 4,304) and validated on the remaining 20% (n = 1,076). Adele integrates 2 jointly fit components: 1) a PK submodel estimating steady-state edoxaban exposure, and 2) a multistate hazard submodel enabling dynamic prediction of subsequent clinical events (eg, probability of CV fatality after a bleeding event; see Figure 1).²³

Figure 1.

Patient States and Transitions in Bayesian Competing-Risk, Multistate Hazard Model

During the time course of chronic therapies, patient characteristics change, that is, they may develop other comorbidities or experience an adverse event, both of which change their risk profile. Arrows denote possible transitions between states. When a subject is in a given state, they may transition to any of the possible next states. In states with multiple possible transitions, each transition represents a competing risk, of which only one can occur at any given time. CV = cardiovascular; GI = gastrointestinal; ICH = intracranial hemorrhage; SEE = systemic embolic event.

Model structure and state transitions

For a detailed description of Adele’s modeling approach and underlying mathematical framework, please see the Supplemental Methods. Adele’s modeling framework starts by including all patients in the randomization state, that is, the patient’s baseline condition at study entry. Subsequently, patients may transition between clinical states based on intercurrent events, with each patient's current state defined by their most recent clinical event/updated clinical history. The model incorporates 9 distinct states, including 3 special categories: the initial randomization state and 2 terminal states (disability and CV fatality) (Figure 1). Possible transitions between states are represented as a directed line, where recurrent events (eg, successive minor bleeding episodes) are shown as “loops” within the same state. Censoring events are treated as observations that terminate follow-up without representing a distinct clinical state.

Bayesian predictive framework

Adele was trained using full Bayesian inference with a state-of-the-art Markov Chain Monte Carlo dynamic Hamiltonian Monte Carlo algorithm based on the No-U-Turn Sampler algorithm²⁴ and the Stan probabilistic programming language.²⁵ R software (R core team 2022) was used for data management, simulations, computation of summary statistics, and graphical analyses. Leveraging Bayesian statistical principles, Adele allows simulation of future clinical event sequences, from which posterior probabilities of event risks and expected event frequencies can be estimated for any patient given their baseline characteristics, drug dosing, and most recent clinical event. This approach enables the generation of counterfactual treatment scenarios, allowing estimation of event rates under alternative dosing regimens not actually assigned to individual patients. Such counterfactual modeling provides quantitative evidence to support personalized dose recommendations by comparing expected outcomes across different therapeutic strategies for each patient's unique profile. Furthermore, it enables dynamic future event rate prediction given an observed or hypothetical clinical event after the start of the study.

ENGAGE AF-TIMI 48 trial study design and patient population

The ENGAGE AF-TIMI 48 (NCT00781391) trial was a 3-arm, randomized, double-blinded trial which compared 2 dose regimens of edoxaban with warfarin. Patients were enrolled from November 19, 2008, through November 22, 2010. Details of the phase 3 ENGAGE AF-TIMI 48 trial (NCT00781391) PK data and clinical events have been published previously.^13,26,27 Briefly, eligible patients were at least 21 years old with AF documented by electrocardiogram within 12 months preceding randomization, a CHADS₂ score of 2 or higher, and planned anticoagulation therapy for the trial duration.¹³ Patients were randomized 1:1:1 to 1 of 3 arms: 1) warfarin, 2) high-dose edoxaban regimen (60 mg daily), or 3) low-dose edoxaban regimen (30 mg daily). Patients in the edoxaban arms had a 50% reduction in dose if 1 or more of the following criteria were met, for example, estimated creatinine clearance (CrCl) <50 mL per minute, body weight ≤60 kg, or concomitant use of the strong cardiac P-glycoprotein inhibitors verapamil, quinidine, or dronedarone.¹³

The study was conducted in accordance with the Declaration of Helsinki and Good Clinical Practice Guidelines of the International Conference on Harmonisation of Technical Requirements for Registration of Pharmaceuticals for Human Use, with approval from independent review boards at each participating site. All patients provided written informed consent before participation.¹³

Population pharmacokinetic submodel

The PK analysis was performed on a subpopulation of 5,380 participants from ENGAGE AF-TIMI 48 who had both pre- and postdose plasma edoxaban measurements within 10 days of the scheduled visit, no quality concerns regarding the PK data, and complete covariate data for both the PK and hazard models. A computationally efficient partial steady-state 1-compartment PK model with linear oral absorption was developed to evaluate the effects of age, sex, CrCl, weight, and dose adjustment on edoxaban PK, in accordance with current best practice.²⁸ The PK submodel included several covariates: age and dose adjustment for k_a (absorption rate constant); age and dose adjustment with sex and CrCl for drug clearance, with dose adjustment plus sex and weight for V₂ (volume of central compartment) (Figure 2).

Figure 2.

Joint Exposure and Hazard Submodels

Adele was trained on a subpopulation of ENGAGE-AF (Effective Anticoagulation with Factor Xa Next Generation in Atrial Fibrillation)-TIMI 48 trial participants selected based on pharmacokinetic data availability and quality. The inclusion criteria required subjects to have both pre- and postdose pharmacokinetic measurements within 10 days of a scheduled visit, and to be free of data quality concerns. Subjects with implausible pharmacokinetic profiles—such as identical pre- and postdose concentrations or anomalous postdose declines—were excluded. Additionally, only subjects with complete covariate data for both the pharmacokinetic and hazard models were retained. Adele was trained on 80% of the population (n = 4,304) and validated on the remaining 20% (n = 1,076). Covariates in the pharmacokinetic submodel included age, sex, weight, creatinine clearance, and dose adjustment, whereas the hazard model incorporated mean drug exposure, dose reduction criteria, country, weight, sex, age, alcohol use, CHADS₂ score, and HAS-BLED (hypertension, abnormal renal or liver function, stroke, bleeding tendency or predisposition, labile INRs, age >65 years, and drugs or alcohol use predisposing to bleeding) score. See Supplemental Methods for additional model details. CrCl = creatinine clearance; HAS-BLED = hypertension, abnormal renal or liver function, stroke, bleeding tendency or predisposition, labile INRs, age >65 years, and drugs or alcohol use predisposing to bleeding.

The final model was qualified using bootstrap and quantitative predictive checks. Average steady-state concentration served as the systemic drug exposure measure incorporated into the hazard submodel.

Hazard submodel

An expansive hazard model was constructed to account for all important competing events and their respective transitions. This submodel included several covariates associated with event hazards: drug exposure, dose reduction criteria, country, weight, sex, age, alcohol use, CHADS₂ score, and HAS-BLED score (Figure 2).

Patient examples

To demonstrate individual patient predictions, we selected 2 patients from ENGAGE AF-TIMI 48 to illustrate the utility of Adele. The patients selected had the following profiles: 1) 80-year-old female with persistent AF, rare alcohol use, body weight of 52 kg, CrCl of 42 mL/min, modified HAS-BLED score of 2, and CHADS₂ score of 3 (patient A); and 2) 72-year-old male with permanent AF, rare alcohol use, body weight of 80 kg, CrCl of 73 mL/min, modified HAS-BLED score of 2, and CHADS₂ score of 4 (patient B). Patient A was randomly assigned to the 30/15 mg dose group and reduced to 15 mg. Patient B was randomly assigned to the 60/30 mg dose group with no dose reduction.

Model assessment and concordance index

For a detailed description of Adele’s modeling approach and underlying mathematical framework, please see the Supplemental Methods. For each subject and each event, we estimate the survival function—the probability that the subject remains event-free over time given the initial state. To assess Adele’s performance, we compare its survival estimates with the Kaplan-Meier (KM) estimator,²⁹ which is stratified by dose amount and dose reduction criteria. The 4 strata are 15 mg, reduced 30 mg, nonreduced 30 mg, and 60 mg. For both Adele and KM, we calculate survival at 3 years based on the randomization initial state. From this, we compute the concordance index (C-index),³⁰ a standard metric in survival analysis that measures predictive accuracy—how well the model ranks subjects by risk, compared to actual event-free duration. The C-index represents the proportion of comparable subject pairs that are concordant: a pair is concordant if the subject who experiences the event first has a lower predicted survival (higher event risk) at 3 years. Pairs are excluded if event order cannot be determined (eg, due to censoring by death or other events). In the present analysis, a positive C-index difference indicates higher predictive accuracy for Adele compared to KM estimators.

Applications of Adele

Predicted cumulative event rates after ICH or IS at 9 months

Estimation of cumulative event rates, that is, the expected number of each event during an interval (0, t), was carried out by generating event sequences using the trained model. Predicted event rates were estimated by generating more than 50,000 event sequences for each patient. This was done for each patient example over the 3-year follow-up in ENGAGE at the randomized dose given conditions at baseline dose at randomization (patient A: 30 mg; patient B: 60 mg), and conditional on an observed hypothetical event at 9 months (intracranial hemorrhage [ICH] or an IS). Changes in predicted event rates following ICH or IS at 9 months do not reflect changes in dosing or other clinical variables.

Patient outcome preferences

Predicted event hazards (rates) were used to identify the optimal edoxaban dose (60 mg, 30 mg, or 15 mg) for 4 example patient outcome preferences: 1) avoid CV death, 2) avoid disability, 3) avoid major bleeding, and 4) avoid any IS/ICH.

Cumulative weighted event rates (3-year) according to hypothetical patient preference

Each patient preference function comprises a set of scores (weights) assigned to each event, according to a hypothetical patient’s preference. Scores range from 0 (no impact) to 1 (worst impact). We define the cumulative weighted event rate as the weighted sum of the expected number of events. The cumulative weighted event rates correspond to the negative expected utility, which was computed for each of the 3 doses. The dose which minimizes the cumulative weighted event rate is the optimal dose according to the hypothetical patient’s preference. The 4 patient preference functions are defined below.

$$AvoidCVDeath\left(e\right)=\{\begin{array}{c}1,ife=CVdeath,\\ 0.9,ife=ICH,\\ 0.7,ife=Ischemicstroke,\\ 0.5,ife=Anynon\text{-}ICHnon\text{-}GImajorbleed,\\ 0.5,ife=AnyMajorGIbleed,\end{array}$$

$$AvoidMajorBleeding\left(e\right)=\{\begin{array}{c}1,ife=Anymajornon\text{-}ICHnon\text{-}GIbleed,\\ 1,ife=ICH,\\ 0.9,ife=MajorGIbleed,\\ 0.5,ife=CVdeath,\\ 0.5,ife=Ischemicstroke,\end{array}$$

$$AvoidIS/ICH\left(e\right)=\{\begin{array}{c}1,ife=Ischemicstroke,\\ 0.9,ife=ICH,\\ 0.7,ife=CVdeath,\\ 0.5,ife=Anynon\text{-}ICHnon\text{-}GImajorbleed,\\ 0.5,ife=MajorGIbleed\end{array}$$

$$AvoidDisability\left(e\right)=\{\begin{array}{c}1,ife=Disability\\ 0,ife=Allotherevents\end{array}$$

CV death includes all adjudicated CV deaths including atherosclerotic vascular disease (excluding coronary), dysrhythmia, systemic embolic event, myocardial infarction, revascularization-related deaths (coronary artery bypass grafting or percutaneous coronary intervention), nonhemorrhagic stroke, congestive heart failure/cardiogenic shock, bleeding, ICH, non-ICH, and pulmonary embolism, and sudden or unwitnessed death. Disability includes all adjudicated events resulting in persistent disability, as defined by the seriousness criteria requiring documentation (eg, an IS or ICH which resulted in persistent disability; see primary ENGAGE AF-TIMI 48 protocol and statistical analysis plan).¹³

Results

Model performance

Adele outperformed the standard dose-group stratified KM estimators across all clinical events, improving 3-year prediction (C-index) in the validation sample (n = 1,076): CV fatality by +12.1%, major gastrointestinal (GI) bleeding by +13.7%, major non-ICH non-GI bleeding by +13.7%, IS by +3.3%, ICH by +4.1%, and disability by +11.8%.

Predicted cumulative event rates

In the illustrative patient examples, predicted cumulative event rates up to 3 years factoring in only baseline information (patient characteristics and the assigned dose) and hypothetical intercurrent events (either IS or ICH at 9 months) are displayed in Figure 3.

Figure 3.

Predicted 3-Year Cumulative Event Rates at Randomization and Following Intracranial Hemorrhage or Ischemic Stroke at 9 Months

Predicted cumulative event rates for 2 patient examples over 3 years (patient A: 80-year-old female; patient B: 72-year-old male). Predicted event rates were estimated by generating more than 50,000 event sequences for each patient and generated using only the patient’s baseline characteristics, edoxaban dose at randomization (patient A: 30 mg; patient B: 60 mg), and conditional on either an intracranial hemorrhage or ischemic stroke at 9 months. Changes in event rates following intracranial hemorrhage or ischemic stroke at 9 months represent updated risk prediction following these intercurrent events, and do not reflect changes in dosing or other clinical variables. CrCl = creatinine clearance; CV = cardiovascular; GI = gastrointestinal; ICH = intracranial hemorrhage; IS = ischemic stroke.

For patient A, the cumulative hazard at randomization was relatively low across all outcome categories, with IS representing the highest baseline risk of approximately 5% over 3 years based on simulations of more than 50,000 potential event sequences. Following ICH, major GI bleeding emerged as the highest risk, reaching approximately 10% cumulative hazard after 3 years. Following IS, disability became the predominant risk, peaking at approximately 8% to 9% cumulative hazard after 3 years.

For patient B, cumulative hazard at randomization showed major GI bleeding and disability as the highest baseline risks, both reaching approximately 6% over 3 years. Following an ICH at 9 months, major GI bleeding and disability emerged as the highest risks, both peaking near ∼9% to 10% over 3 years. Following IS at 9 months, disability, major GI bleeding, and major non-ICH non-GI bleeding became the predominant risks, with each reaching approximately 6.5% to 7% after 3 years.

Minimizing cumulative weighted event rates (3-Year) according to patient preference

Avoiding CV death

The avoid CV death weighted event rates varied across doses and between both patient examples, denoting differences in patient baseline characteristics. In patient A, when CV death is ranked as the most severe outcome, the predicted cumulative 3-year event rates for edoxaban 15 mg, 30 mg, and 60 mg were 10.2%, 11.5%, and 13.7%, respectively. In patient B, the corresponding rates were 14.0%, 13.1%, and 11.9% for edoxaban 15 mg, 30 mg, and 60 mg, respectively (Figure 4, Panel 1A & B).

Figure 4.

Expected Weighted Cumulative 3-Year Event Rates According to Hypothetical Patient Preference Across Edoxaban Doses

Predicted cumulative 3-year event rates after weighting events according to a hypothetical patient preference, placing the greatest importance on avoiding cardiovascular death (1A & B), avoiding disability (2A & B), and avoiding major bleeding or ischemic stroke/intracranial hemorrhage (3A & B), across 3 hypothetical dosing options (15 mg, 30 mg, 60 mg, ie, current clinically available edoxaban doses) in patient A and B. Each colored segment represents the (weighted) contribution of specific event subtypes. The cumulative totals (∑) for each dose are displayed above each bar. Event rates shown are hypothetical and represent the rate of events expected to occur within a 3-year period based on the patient risk profile at baseline and incorporating expected intercurrent event frequencies and subsequent risks. See Methods for specific weighting assignments. CV = cardiovascular; GI = gastrointestinal; ICH = intracranial hemorrhage; IS = ischemic stroke.

Avoiding disability

Predicted disability rates varied across doses and between patients (Figure 4, 2A & B). In patient A, the predicted 3-year disability rates for edoxaban 15 mg, 30 mg, and 60 mg were 4.9%, 4.4%, and 5.6%, respectively. For patient B, the corresponding rates were 1.9%, 2.0%, and 2.5% for edoxaban 15 mg, 30 mg, and 60 mg, respectively.

Avoiding major bleeding or IS/ICH

Alternative rankings were applied to depict potential differences in hypothetical patient preferences to either avoid major bleeding (any ICH or other non-ICH major bleeding, except major GI bleeding) for patient A, and avoid IS/ICH for patient B. In patient A, the avoid major bleeding weighted predicted 3-year event rates for edoxaban 15 mg, 30 mg, and 60 mg were 11.6%, 13.7%, and 18.0%, respectively (Figure 4, Panel 3A). In patient B, the avoided IS/ICH weighted rates were 14.5%, 13.9%, and 12.3%, respectively.

Discussion

Adele advances ML-guided, longitudinally adaptive modeling that supports personalized anticoagulation decisions by integrating PK modeling with multistate event prediction. This approach enables dynamic, patient-centered risk estimation and facilitates shared decision-making, representing a significant evolution beyond static, population-level dosing recommendations (Central Illustration).

Central Illustration.

Bayesian Machine Learning Model Guiding Iterative, Personalized Decision-Making in Anticoagulant Treatment

Image to demonstrate use case of the Adele Bayesian machine learning model in an interactive patient application displayed on a tablet device. Adele provided dynamic predictions for 3-year events and optimal edoxaban dose based on hypothetical patient inputs, thus providing an adaptive patient-centric tool based on patient preference. CrCl = creatinine clearance; CV = cardiovascular; ENGAGE AF-TIMI 48 = Effective Anticoagulation with Factor Xa Next Generation in Atrial Fibrillation–TIMI 48; GI = gastrointestinal; ICH = intracranial hemorrhage; IS = ischemic stroke; ML = machine learning; PK = pharmacokinetic.

Adele was superior to static KM estimates in predicting CV fatality, major GI bleeding, major non-ICH non-GI bleeding, IS, ICH, and disability, with improvements ranging from +3.3% to +13.7% in C-index. In the patient examples, we demonstrated several key advances: first, the model's dynamic risk prediction capabilities, where subsequent event probabilities shifted following intercurrent events like ICH or IS; second, the impact of individual patient characteristics on optimal dosing, as evidenced by patient A (avoid major bleeding) showing optimal outcomes at lower doses (15-30 mg) regardless of preference ranking, whereas patient B (avoid IS/ICH) achieved optimal outcomes at higher doses (60 mg) for IS/ICH and CV death prevention and overall low disability risk across doses. These findings highlight that individual risk profiles and preferences can influence optimal dosing strategies. Adele’s ability to provide conditional probabilities of subsequent events, demonstrating how cumulative hazard increases differentially following various/multiple events, offers clinicians and patients unprecedented insight into dynamic risk trajectories that static risk scores cannot capture.

ENGAGE PK data and analyses as the foundation for adele

The high-quality PK data from the phase 3 ENGAGE AF–TIMI 48 trial and the results from their subsequent analyses set foundation for Adele. ENGAGE yielded the first comprehensive analyses of any DOAC to integrate validated drug concentration measurements (using liquid chromatography-tandem mass spectrometry with a lower limit of quantification of 0.764 ng/mL), biologically relevant pharmacodynamic markers (endogenous factor Xa activity), and adjudicated clinical efficacy and safety outcomes.^26,27 Specifically, Ruff et al demonstrated that the range of edoxaban doses 15 to 60 mg resulted in a 2-fold to 3-fold gradient of mean trough drug exposure (16.0-48.5 ng/mL in 6,780 patients with data available) and mean trough antifactor Xa activity (0.35-0.85 IU/mL in 2,865 patients). Dose reduction decreased edoxaban exposure by 29% and 35% and antifactor Xa activity by 20% and 25% across the high- and low-dose regimens. Despite these lower levels, dose reduction maintained edoxaban's efficacy against the occurrence of thromboembolic stroke compared with warfarin, while providing significantly greater protection against major bleeding.²⁶

Yin et al subsequently advanced this work by characterizing endogenous factor Xa activity as a more biologically relevant pharmacodynamic marker, demonstrating that inhibition of endogenous factor Xa activity is saturated above 440 ng/mL edoxaban concentration. Further, that the degree of inhibition—influenced by both edoxaban dose and patient characteristics (eg, age, body weight, sex)—is directly associated with both antithrombotic benefit and major bleeding risk, with model predictions enabling quantification of personalized risk profiles across the therapeutic range.²⁷

Together, these 2 analyses provided the foundation for the Bayesian ML framework employed by Adele, which shall inform dose selection in future trials, guide patient monitoring, and optimize dynamic patient-centric oral anticoagulation therapy.

Heterogeneity in patient preferences drives need for personalized approaches

Current evidence demonstrates remarkable heterogeneity in patient preferences for anticoagulation decisions, potentially confounding the effectiveness of standardized clinical decision-making approaches. Discrete choice experiments reveal that patients prioritize stroke prevention and bleeding risk differently, with preference weights varying considerably by demographics and prior experiences.^31,32 For example, in a sample of US patients with CV disease, patients valued a 1% increased risk of a fatal bleeding event the same as a 2% increase in nonfatal myocardial infarction, a 3% increase in nonfatal stroke, a 3% increase in CV death, a 6% increase in major bleeding, and a 16% increase in minor bleeding.³¹ Furthermore, prior stroke or myocardial infarction was associated with a larger negative ranking for these outcomes (highlighting the preference to avoid recurrent events).³¹ In a survey of 172 hospitalized patients with AF in Canada, LaHaye et al.³³ found striking variability in risk tolerance: although 12% were completely medication averse and refused anticoagulation even if it eliminated stroke risk, others were willing to accept up to 4 major bleeding events to prevent a single stroke.³³ The study further identified distinct patient subgroups—42% were risk averse, 15% risk tolerant, and the remainder balanced risks and benefits—demonstrating that preferences for stroke prevention vs bleeding avoidance are highly individualized. A separate cross-sectional survey of both outpatients and inpatients with AF from 2 public hospitals in China (n = 506) showed systematic variations in preference weights: stroke/systemic embolism risk, major bleeding risk, and acute myocardial infarction risk, but with substantial individual variation around these means.³²

The hypothetical “patient preference” weights in our model are intended to reflect these real-world patterns and can be adapted to different populations or individual values, as suggested by published studies.31, 32, 33 This flexibility allows Adele to accommodate patients who may prioritize stroke prevention or bleeding avoidance to varying degrees.

Efforts to accommodate these preferences through the implementation of shared decision-making (ie, patient/care provider working together to align on medical care) remains disappointingly low, despite the American College of Cardiology, the American Heart Association, and the European Society of Cardiology medical guideline recommendations.^4,12 Recent evidence from the Survey of Patient Knowledge and Personal Priorities for Treatment (SATELLITE), a substudy within the national Outcomes Registry for Better Informed Treatment of Atrial Fibrillation II (ORBITII) registry reveals that only 26% of patients reported that their current stroke prevention treatment strategy resulted from shared decision-making with their healthcare provider.^34,35 Perhaps more concerningly, approximately half of patients reported that decisions about oral anticoagulation and rhythm control were made entirely by their healthcare providers, without any patient input. This implementation gap persists despite strong evidence that shared decision-making reduces costs and improves outcomes.³⁶ The primary barrier cited is a lack of adequate tools and technologies to facilitate meaningful patient engagement in complex risk-benefit discussions.³⁴

Adele directly addresses this deficit as a tool designed specifically for shared decision-making in anticoagulation therapy, by translating individual patient preferences into quantitative preference functions that can guide dose selection. By enabling clinicians to model “what-if” scenarios (counterfactual) based on patient-specific values, Adele supports genuine shared decision-making where patients can visualize how their preferences translate into concrete treatment recommendations.

Static bleeding risk prediction fails over time

Current bleeding risk assessment tools lack detailed risk stratification, particularly in older adults in whom age increasingly dominates scoring, inappropriate focus on all bleeding events rather than major/fatal bleeds, and inability to determine the timing of imminent bleeding risk. These tools also struggle with competing events and offsetting risk factors (eg, treatment duration potentially reducing risk whereas aging increases it), leading to critical time-dependent decay of predictive accuracy. Post hoc analyses from major clinical trials reveal that established bleeding risk scores such as HAS-BLED, ORBIT-AF, Anticoagulation and Risk Factors in Atrial Fibrillation (ATRIA) and HEMORR₂HAGES (a validated scoring system for major bleeding risk, particularly in AF patients, of the following key factors: hepatic or renal disease, ethanol abuse, malignancy, age >75 years, reduced platelet count or function [eg, aspirin use], history of re-bleeding, hypertension [uncontrolled], anemia, genetic factors, excessive fall risk, history of stroke) provide predictive benefit only within the first 3 years of treatment, with no predictive benefit beyond 5 years.³⁷ This decline over time reflects the dynamic nature of bleeding risk, which evolves with aging, incident comorbidities, and concurrent therapies—factors that static scores cannot accommodate.

Evidence from clinical practice and registry compounds these concerns. Patients who experience major bleeding incur a 2–3-fold higher risk of subsequent bleeding events, yet current management approaches lack standardized protocols for postbleeding anticoagulation decisions.^38,39 Registry data show that, following a major bleeding event, only 54% of patients receive optimal anticoagulation (defined as either vitamin K antagonist for an international normalized ratio of 2-3 or appropriately dosed DOAC according to the product label), whereas in 26% the regimen was suboptimal (defined as anticoagulation at lower dose levels than label guidance or antiplatelet therapy alone), highlighting inconsistent clinical implementation of risk-benefit assessments.^38,39 In a meta-analysis of 16 randomized controlled trials or observational studies comparing off-label vs on-label oral anticoagulant dosing for patients with AF (N = 130,609), 29% (9.4% to 50.0%) and 4% (3.4% to 50.0%) were either underdosed or overdosed off-label.⁴⁰ In the largest pooled analysis conducted to date (N = 223,057 patients with AF from 22 studies), compared with on-label doses, off-label, underdosed DOACs were associated with a 26% higher risk of all-cause death (HR: 1.26; 95% CI: 1.09 to 1.43) but not an increased risk of any stroke (HR: 1.03; 95% CI: 0.88-1.17) whereas off-label DOAC overdosing increased the risk of all-cause death (HR: 1.19; 95% CI: 1.03-1.35), any stroke (HR: 1.17; 95% CI: 1.04-1.31), and major bleeding (HR: 1.18; 95% CI: 1.05-1.31).⁴¹

Adele’s Bayesian framework addresses the limitations of suboptimal dosing, as it updates risk continuously based on intercurrent events and evolving patient characteristics. The model’s ability to provide conditional probabilities following any events—such as the demonstrated differential risk trajectories after ICH vs IS—enables clinicians to make evidence-based decisions in those clinical scenarios where current tools fall short.

The development of ML-driven personalized medicine tools like Adele represents a potential advancement in therapeutic decision-making which may inform clinical practice and regulatory frameworks. The U.S. Food and Drug Administration has emphasized the importance of dosing to optimize benefit vs risk in individual patients.^6,7 Adele exemplifies the type of dynamic, evidence-based tool that regulatory agencies are encouraging—one that combines mechanistic understanding (PK) with evidence from clinical practice (clinical outcomes) and patient preferences.

Study Limitations

Adele was trained exclusively on data from the randomized ENGAGE AF-TIMI 48 clinical trial and is directly applicable only to edoxaban. However, the framework is general and extension to other DOACs would only require retraining. While 2 illustrative patient examples are presented to demonstrate the model’s capabilities, future work will include additional patient profiles and preference functions. Although preference functions were evaluated across the 3 edoxaban doses (60 mg, 30 mg, 15 mg) the model could, in principle, identify optimal doses between these values. However, a key limitation is that clinical application is constrained by the 3 marketed tablet doses (note: approved edoxaban doses vary throughout the world; eg, edoxaban 15 mg is approved in Japan, China, and other Asian countries, but not in Europe or the United States; prescribing practices should adhere to the approved local label); thus, even if the model suggests an ideal dose that falls between available doses, only the approved doses can be prescribed in practice. Consideration should be given to adaptive dosing in the future. Adele requires prospective validation in diverse patient populations and real-world clinical settings to address potential biases arising from nonrandom treatment assignment before clinical implementation.

Conclusions

Adele is a novel Bayesian ML framework that provides dynamic, personalized recommendations for dosing of the oral anticoagulant edoxaban. By combining mechanistic PK modeling with probabilistic event forecasting and user-defined patient preference functions, it enables data-driven, patient-centered care. This paradigm supports iterative, shared decision-making, offering a path toward potentially safer and more effective anticoagulation therapy and applicable to other chronic treatments.

Perspectives.

COMPETENCY IN MEDICAL KNOWLEDGE: Adele is a state-of-the-art Bayesian framework, leveraging clinical factors and PK data, to enable longitudinally adaptive, patient-specific, preference-weighted predictions. Compared to standard Kaplan-Meier estimators, Adele improved 3-year prediction accuracy for all clinical outcomes. Adele's ability to update risk estimates after intercurrent events and to incorporate patient outcome preferences supports a framework for more precise, patient-centered anticoagulation management.

TRANSLATIONAL OUTLOOK: The adoption of dynamic tools like Adele could transform not only anticoagulation management, but also the care of other chronic diseases. Successful clinical translation of Adele’s framework will require external validation in real-world datasets and evaluation of its impact on clinical outcomes, ideally through randomized clinical trials. Collaboration with regulatory agencies will be essential for acceptance and integration into clinical practice. Additionally, developing user-friendly interfaces to support clinician and patient engagement—and to facilitate shared decision-making—will be key for effective implementation in real-world clinical workflows.

Funding support and author disclosures

This study was funded by Daiichi Sankyo, Inc, Basking Ridge, New Jersey, USA. Dr Gibson has received research grant support from Johnson & Johnson and Bayer. Ms Buros-Novik and Mr Novik are employees of Generable Inc, New York, New York, USA. Mr Timonen has received consulting fees from Generable Inc, New York, New York, USA, for work related to the research presented in this paper. Dr Braunwald has received research grants from AstraZeneca, Daiichi Sankyo, Merck, and Novartis; and consulting fees from Amgen, Boehringer Ingelheim, Bristol Myers Squibb, Cardurion, Edgewise, IMMEDIATE, and Verve. Dr Antman has received research grants from Daiichi Sankyo. Dr Fronk is an employee of Daiichi Sankyo Europe GmbH, Munich, Germany, and may have stock or stock options. Drs Chen, Clasen, and Unverdorben are employees of Daiichi Sankyo Inc, Basking Ridge, New Jersey, USA, and may have stock or stock options. Dr Giugliano has received research grants from Anthos Therapeutics, Daiichi Sankyo, and Novartis; honoraria from Big Sky Cardiology, Daiichi-Sankyo, Dr Reddy’s Laboratories, Medical Education Resources (MER), Menarini, SAJA Pharmaceuticals, Shanghai Medical Technology, and SUMMEET; and consulting fees from Artivion, Inc, Celecor, Daiichi Sankyo, Inari, Novartis, Perosphere, PhaseBio, Samsung, Sanofi, SFJ Pharmaceuticals, and Thrombosis Research Institute.