Predict PaO2/FiO2 from non-laboratory data

The goal of this repository is to predict the PaO2/FiO2 ratio from non-laboratory data

Review of bilan2015comparison

Primary application was predicting pao2/fio2 from spo2/fio2 in ards

All patients had to have a previous PaO2/FiO2 ratio <= 300 before enrollment.

data collection:

• based on various studies
• 24 hours BEFORE ENROLLMENT:
  • ventilator parameters, study hospital, volume of fluid administered in the 24 hours before enrollment
• AT ENROLLMENT:
  • Age, sex, BMI, mean arterial pressure, and use of vasopressors at enrollment
    • any dose of norepinephrine, epinephrine, dopamine, phenylephrine, or vasopressin
• DAY OF ENROLLMENT:
  • ABG closest to 8 am on the day of enrollment
    • serum bilirubin, FIO 2 and SpO 2 at time of ABG

data preprocessing

• excluded patients who did not have FIO2, PaO2, and SpO2 recorded from an ABG
• adjusted PaO2/FIO2 ratios at the Denver and Utah sites (altitude ~1500m) by the ratio of local to sea level barometric pressure (0.836 in Denver, 0.845 in Utah).

equations used

• non-linear imputation based on the Severinghaus equation
  • Oxygen Saturation = (23,400 (po^3 + 150 po^2)-1 + 1)-1
  • inverting this equation is fun I promise
• log-linear imputation based on the Pandharipande equation
  • Log(PF) = 0.48 + 0.78 x Log(SF)
• linear imputation based on the Rice equation
  • S/F = 64 + 0.84 * (P/F)

data analysis

• correlation between PaFi measured, PaFi imputed
  • once for all patients
  • once for patients w/ SpO2 <= 96
• RMSE of measured/imputed PaFi
• Built a regression with Imputed PaO2 + other features, outcome measured PaO2
• For the PaO2/FIO2 thresholds that were used to define mortality strata in the Berlin ARDS definition, we calculated the imputed PaO2/FIO2 that was associated with the same mortality as the measured PaO2/FIO2 threshold

results

• PaO2/FiO2 - all patients (N=1184)

  • correlations
    • 0.84 - non-linear
    • 0.73 - log-linear
    • 0.73 - linear
  • RMSE
    • p=0.92 non-linear vs log-linear
    • p<0.001 non-linear vs linear

• PaO2 - all patients

  • correlations
    • 0.72 for non-linear imputation
    • 0.30 for log-linear
    • 0.13 for linear
  • RMSE
    • p<0.001 non-linear vs log-linear
    • p<0.001 non-linear vs linear

• Patients with SpO2 <= 96%

  • PaO2/FiO2 - correlations
    • 0.90 - non-linear
    • 0.88 - log-linear
    • 0.88 - linear
  • RMSE - PaO2/FiO2
    • 51.7 - non-linear
    • 52.0 - log-linear
    • 66.4 - linear
  • RMSE - PaO2
    • 28.6 - non-linear
    • 32.2 - log-linear
    • 46.4 - linear
  • All RMSE p-values < 0.0001

• PaO2 - patients with SpO2 <= 96% (N=707)

  • correlations
    • 0.72 for non-linear imputation
    • 0.13 for linear
    • 0.30 for log-linear
  • RMSE
    • p<0.001 non-linear vs linear
    • p<0.001 non-linear vs log-linear

Confusion matrix for PaO2/FiO2 > 200 (0 = lower, 1 = higher i.e. severe ARDS)

 | 0   | 1 (Imputed)

--- | --- | --- 0 | 764 (65%) | 101 (9%) 1 (true) | 70 (6%) | 249 (20%)

Concordance was not associated with mortality after controlling for age, PEEP, and APACHE III score. "The sickness of the patient is not related to the concordance of this test".