Regenerating Evidence from Landmark Trials in ARDS Using Structural Causal Models on Electronic Health Record

Rationale: Randomized Controlled Trials (RCTs) are considered the Gold Standard to establish causality in medicine. However, RCT’s can be time consuming, expensive and often unethical. We present how machine learning models, more specifically structural causal model (SCM) can be used for data driven clinical evidence generation. We validated our model by replicating the results of three well known RCT’s with treatment interventions for acute respiratory distress syndromes (ARDS). Methods: The Medical Information Mart for Intensive Care (MIMIC III) database is a freely available source of critical care EHRs for 53,423 adult admissions and 7,870 neonates. MIMIC III database includes patient demographics, time-stamped measurements of physiological vitals, administration of medications, notes charted by caregivers, as well as diagnoses encoded as ICD-9. We used MIMIC III database to derive our study cohorts to create a structural causal model (SCM). The algorithm was developed in two stages: i) structure learning algorithms and clinical consultation to learn the causal structure, and, ii) adjustment formula to derive the equation for counterfactual intervention. In this paper, we are focusing on reproducing the result of three landmark trials in treating ARDS patients: ARMA, ALVEOLI, and ACURASYS trials. We examined the medical records of patients who had ARDS onset within 48 hours from admission as selected RCTs. For each cohort the following variables were collected: sex, age, weight, height, SOFA score, APACHE-III score, PaO2:FiO2, tidal volume, plateau pressure, peak inspiratory pressure, positive end-expiratory pressure, fraction of inspired oxygen, respiratory rate, minute volume and arterial pH, PaO2, PaCO2 and SPO2. Results: Regenerating the ARMA trial with MIMIC III database included 1411 patients with the inclusion criteria (Berlin score <300). The survival rate was found to be 0.69 in the treatment group in our observational study compared to 0.69 in the trial. However discordant results were seen in control group with a survival rate of 0.9 in our study vs 0.6 in the trial. This can be explained by lower APACHE score in our cohort control group in comparison to the trial. Our observational data for ALVEOLI trial, showed similar survival rates in our study compared to the trial in treatment arm 0.73 vs 0.75 respectively and control arm 0.65 vs 0.73. We saw survival rates of our observational data in comparison to ACURASYS trial to be very similar in treatment arm 0.66 vs 0.68 and control arm 0.57 vs 0.58 (p <0.01 for each result). Conclusion: Clinical evidence can be reliably derived from existing observational data using structural causal models. However, one of the major limitations is lack of methods to stratify data groups. There are also limitations of selection bias and heterogeneous population. These issues can be addressed using methods to remove selection bias and transportability in structural causal model.