EveryQuery: Zero-Shot Clinical Prediction via Task-Conditioned Pretraining over Electronic Health Records
Revolutionizing EHR Prediction with EveryQuery: Faster, More Accurate Zero-Shot AI
EveryQuery is a groundbreaking foundation model that addresses critical limitations of current autoregressive (AR) models for Electronic Health Records (EHRs). By introducing task-conditioned pre-training, EveryQuery enables direct, prompted zero-shot clinical prediction, overcoming the computational expense, noise, and lack of promptability inherent in simulation-based AR approaches. This innovation significantly accelerates inference and improves prediction accuracy across diverse clinical tasks, especially for rare events.
Transforming Clinical AI Capabilities
EveryQuery offers a paradigm shift in how AI models interact with EHR data, delivering substantial gains in efficiency and reliability, which are crucial for real-world clinical applications.
15%
Max AUC Improvement
10,000x
Max Inference Speedup
84.6%
Max Win Rate vs. AR
149M
Model Parameters
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Efficiency & Speed
EveryQuery dramatically reduces the computational burden of zero-shot EHR prediction, making advanced AI more accessible and practical for real-time clinical use.
10,000x
Faster Inference Speed
EveryQuery achieves up to 10,000 times faster inference speeds compared to autoregressive baselines by directly predicting outcomes rather than simulating future trajectories. This translates into significant cost savings and enables real-time decision support.
Single
Forward Pass
Unlike AR models requiring thousands of GPU hours and multiple trajectory simulations, EveryQuery makes a prediction in a single deterministic forward pass, fundamentally streamlining the inference process.
Accuracy & Reliability
EveryQuery consistently outperforms autoregressive models in prediction accuracy, especially for critical rare events, providing more dependable insights.
84.6%
Win Rate on MIMIC-IV
Across three datasets, EveryQuery achieves high win rates against AR baselines, peaking at 84.6% on MIMIC-IV, demonstrating superior zero-shot prediction performance.
15.4%
Max AUROC Improvement
EveryQuery shows a mean AUROC improvement of +15.4% on MIMIC-IV, consistently outperforming AR models in aggregate per-task AUROC.
Prevalence-Invariant
Performance on Rare Events
EveryQuery's accuracy is less affected by event prevalence, making it particularly effective for predicting rare but clinically important outcomes, a significant weakness of AR models.
Task Conditioning & Promptability
EveryQuery introduces promptable, task-conditioned representations, allowing for flexible and specific queries directly at inference time.
Enterprise Process Flow
Patient EHR History (x)
→
Structured Task Query (q = c, ∆t)
→
EveryQuery Model (Task-Conditioned Pre-training)
→
Direct Prediction (p(y=1 | x,q))
→
Zero-Shot Clinical Inference
EveryQuery takes both patient history and a structured task query as input, directly predicting the outcome without simulation. This task-conditioned approach enables explicit promptability.
Structured
Embedding Space
The model produces a richly structured embedding space where representations are organized by query code, duration, and patient identity, confirming that task prompts directly inform the model’s learned geometry.
Monotonic
Prediction Horizons
EveryQuery demonstrates internal consistency, with predicted probabilities largely non-decreasing as task duration increases (92.8-96.1% across datasets), indicating a coherent understanding of time.
Generalizability & Limitations
While highly generalizable, the model acknowledges current limitations in query language complexity and generalization to unseen codes.
Composite Queries: 30-Day Readmission
EveryQuery successfully tackles composite tasks, such as 30-day hospital readmission (a disjunction over 70 admission codes), by aggregating individual code predictions. It achieved an AUROC of ~0.65, matching the autoregressive baseline, indicating its extensibility beyond single-code queries. This demonstrates robust performance even when the query language requires creative aggregation.
Adversarial Code Withholding
When 100 randomly chosen codes were withheld from pre-training tasks, EveryQuery still achieved a 74.7% win rate on MIMIC-IV (compared to 84.6% standard). However, performance dropped significantly on smaller datasets like NWICU (34.4% win rate). This highlights both the model's capacity for generalization in code embeddings and the risk of overfitting on task sampling for smaller datasets, pointing to areas for future research.
SQL/ACES
Future Query Languages
Future work aims to extend EveryQuery’s query language to support more complex tasks, potentially incorporating SQL queries or ACES configuration files to cover a wider variety of clinical prediction scenarios.
Calculate Your Potential ROI
Estimate the potential efficiency gains and cost savings for your enterprise by leveraging EveryQuery's advanced clinical prediction capabilities. Adjust the parameters below to reflect your organization's scale and operational costs.
Annual Savings
$0
Hours Reclaimed Annually
0
EveryQuery Implementation Roadmap
A phased approach to integrating EveryQuery into your clinical workflows, designed for seamless adoption and measurable impact. Our experts guide you through each step.
Phase 1: Data Preparation & Model Training
Initial setup, including EHR data preprocessing into the MEDS format and task-conditioned pre-training of EveryQuery on your specific datasets. This ensures the model is optimized for your clinical environment.
Phase 2: Task Definition & Integration
Collaborative definition of target clinical prediction tasks using EveryQuery’s structured query language. Integration with existing clinical decision support systems and validation of zero-shot inference capabilities.
Phase 3: Performance Validation & Deployment
Rigorous evaluation against baselines and clinical benchmarks, focusing on AUC improvements and inference speed gains. Deployment into a secure, monitored environment for real-time clinical prediction and continuous improvement.
Ready to Transform Your Clinical Predictions?
EveryQuery delivers unparalleled speed and accuracy for zero-shot clinical AI. Discover how our task-conditioned pre-training can empower your organization with direct, reliable, and promptable insights from EHR data. Book a session to explore a tailored implementation plan.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat