Structural causal model with expert augmented knowledge to estimate the effect of oxygen therapy on mortality in the ICU

Abstract

Recent advances in causal inference techniques, more specifically, in the theory of structural causal models, provide the framework for identifying causal effects from observational data in cases where the causal graph is identifiable, i.e., the data generation mechanism can be recovered from the joint distribution. However, no such studies have been performed to demonstrate this concept with a clinical example. We present a complete framework to estimate the causal effects from observational data by augmenting expert knowledge in the model development phase and with a practical clinical application. Our clinical application entails a timely and essential research question, the effect of oxygen therapy intervention in the intensive care unit (ICU). The result of this project is helpful in a variety of disease conditions, including severe acute respiratory syndrome coronavirus-2 (SARS-CoV-2) patients in the ICU. We used data from the MIMIC-III database, a widely used health care database in the machine learning community with 58,976 admissions from an ICU in Boston, MA, to estimate the oxygen therapy effect on morality. We also identified the model's covariate-specific effect on oxygen therapy for more personalized intervention.

Keywords: Causal inference; Critical care; Expert augmented knowledge; Oxygen therapy; Structural causal model.

Figure 8:

Causal Graph $G_{\overline{\mathit{\text{oxygenation,pao}}2,vt,\mathit{\text{minVentVol}}}}$.

Figure 9:

Causal Graph $G_{\overline{U-\left\{\mathit{\text{age}}\right\}}},G_{\overline{U-\left\{\mathit{\text{gender}}\right\}}},G_{\overline{U-{\mathit{\text{SpO}}}_2}},G_{\overline{U-\left\{ph\right\}}},G_{\overline{U-\left\{\mathit{\text{ards}}\right\}}},G_{\overline{U-\left\{\mathit{\text{bmi}}\right\}}}$.

Figure 10:

Causal Graph $G_{\overline{U-\left\{\mathit{\text{peep}}\right\}}}$.

Figure 11:

Causal Graph $G_{\underset{\_}{\mathit{\text{age}}}}$.

Figure 12:

Causal Graph $G_{\overline{U-\left\{\mathit{\text{copd}}\right\}}}$.

Figure 13:

Causal Graph $G_{\underset{\_}{\mathit{\text{age,spo}}2}}$.

Figure 14:

Causal Graph $G_{\overline{U-\left\{\mathit{\text{fio}}2\right\}}}$.

Figure 15:

Causal Graph $G_{\underset{\_}{\mathit{\text{peep,copd,age,spo}}2}}$.

Figure 16:

Causal Graph $G_{\overline{U-\left\{\mathit{\text{hemoglobin}}\right\}}}$.

Figure 17:

Causal Graph $G_{\underset{\_}{\mathit{\text{gender,spo2}}}}$.

Figure 18:

Causal Graph $G_{\overline{U-\left\{\mathit{\text{trauma}}\right\}}}$.

Figure 19:

Causal Graph $G_{\underset{\_}{\mathit{\text{peep,age}}}}$.

Figure 20:

Causal Graph $G_{\overline{U-\left\{\mathit{\text{smoker}}\right\}}}$.

Figure 21:

Causal Graph $G_{\overline{U-\left\{\mathit{\text{paco}}2\right\}}}$.

Figure 22:

Causal Graph $G_{\underset{\_}{\mathit{\text{copd,spo}}2,\mathit{\text{ph,smoker,age,ards,bmi}}}}$.

Figure 23:

Causal Graph $G_{\overline{U-\left\{\mathit{\text{apsiii}}\right\}}}$.

Figure 24:

Causal Graph $G_{\underset{\_}{S-\left\{\mathit{\text{apsiii}}\right\}}}$.

Figure 25:

Causal Graph $G_{\overline{U-\left\{\mathit{\text{peakAirPressure}}\right\}}}$.

Figure 26:

Causal Graph $G_{\underset{\_}{S}}$.

Figure 27:

Causal Graph $G_{\overline{U-\left\{\mathit{\text{lactate}}\right\}}}$.

Figure 28:

Causal Graph $G_{\underset{\_}{I}}$.

Figure 29:

Causal Graph $G_{\overline{U-\left\{\mathit{\text{death}}\right\}}}$.

Figure 30:

Causal Graph $G_{\underset{\_}{H}}$.

Figure 31:

Causal Graph $G_{\overline{U-\left\{\mathit{\text{sofa}}\right\}}}$.

Figure 32:

Causal Graph $G_{\underset{\_}{G}}$.

Figure 1:

Schematic diagram for the causal estimation framework. The framework allows domain expertise to be encoded in the SCM.

Figure 2:

Inclusion-exclusion criteria for (a) pilot RCT and (b) for the observational study with the MIMIC database.

Figure 3:

Pooled frequency histogram of the percentage of time spent at each SpO₂ level in both oxygenation groups for OT-RCT (Panwar et al., 2016) and our study, SCM-VRCT. SpO₂ is oxygen saturation as measured by pulse oximetry.

Figure 4:

Causal Graph G based on the majority voting of SLAs. oxygenation is the treatment variable, and death is the outcome variable. trauma, surgery and medical nodes signify the diagnosis type for individual patients.

Figure 5:

Causal Graph G after incorporating domain knowledge through expert reviews. oxygenation is the treatment variable, and death is the outcome variable. trauma, surgery and medical nodes signify the diagnosis type for individual patients.

Figure 6:

Estimation of queries on the causal graph for both observational and experimental dataset. (a) P_oxygenation(death), (b) P_SpO2(death) and (c) P(death|oxygenation)

Figure 7:

Estimation of queries on the causal graph for both observational and experimental dataset. (a) P_oxygenation(death|age), (b) P_oxygenation(death|apsiii, sofa) and (c) P(death|oxygenation)