Understand why patient trajectories unfold, and what would change them
Healthcare partners we work with
From correlation to cause, at the bedside and across populations
Causal inference
Time-series expertise
Hybrid digital twins
Privacy first
ICU policy counterfactual simulator
A working research prototype built on 6,460 ICU sepsis stays from MIMIC-IV, recorded in 8-hour steps. Clinicians describe a patient or a cohort and write a treatment policy in plain English. Minutes later, they see what the model expects under that policy compared with the same patients with the drug withheld, next to the trials that tested something similar.
- Digital twins and cohorts
Find patients like a trial patient, or define a cohort by immune state, comorbidities, lactate and MAP.
- Trial-style comparison
The policy vs. the same patients with the drug withheld, from the same starting point and random draws.
- Doubly robust check
AIPW estimates with a positivity verdict that says when the data can't answer.
- Robustness
Every apparent benefit is rerun with a second model variant.
- Evidence
A 17-study library (ADRENAL, APROCCHSS, CENSER, NICE-SUGAR and more), each with a PubMed link.
- Analyst agent
A local LLM writes causal insights from facts computed in code; a clinician manual shows how to review them.
- What it found, honestly
Model estimates on retrospective data, not clinical recommendations.
- Early antibiotics: benefit in both models, confirmed by the doubly robust estimate.
- Hydrocortisone for refractory shock in hyperinflammatory patients: the first steroid policy both models agree on. It mirrors how ADRENAL and APROCCHSS enrolled, but needs prospective confirmation.
- Vasopressors: the model reads them as harmful, a known confounding-by-indication bias. The positivity check flags it, so it is not reported as a finding.
Providers, payers and research partners
PeaceHealth & NIH
Daily steps rose from about 3,000 to 8,500 in a randomized trial; 40+ papers published.
Challenge
Use social networks to spread healthy behaviours.
Approach
A personal health social network with wearables. Deep learning on sensor and social data, and causal models of how behaviours spread through the social graph.
Result
Statistically significant weight loss in an RCT, and daily activity rising from about 3,000 to 8,500 steps.
Humana
Deployed in operations for millions of members; a ComputerWorld best-of-breed award.
Challenge
Move millions of members from reactive care to proactive, integrated health management.
Approach
An integrated health management platform with time-series feature engineering and deep learning to flag high-risk transitions, and causal models to choose the care interventions with the best long-term outcomes.
Result
Deployed in operations.
Mcure Biosciences
A hybrid ICU digital twin for an upcoming clinical trial.
Challenge
High-stakes ICU treatment pathways, with fragmented real-world evidence.
Approach
A hybrid digital twin using multi-headed transformers, time-series clustering and hidden Markov models to account for patient heterogeneity. Causal AI estimates direct intervention effects and simulates counterfactual treatments.
Status
Working with Mcure's clinical and data scientists to deploy digital twins in an upcoming trial.
Simulate your next trial before you run it
Bring a population and a treatment question. We'll show what the digital twins and the evidence say.
Book a working session