Establishment and evaluation of an image-based clinical prediction model for invasive pulmonary fungal infection

Objective To establish and evaluate a prediction model for invasive pulmonary fungal infection based on clinical images in order to provide reference for early identification of the disease. Methods A case-control trial was designed and conducted on 187 hospitalized patients with invasive pulmonary fungal infection and 190 patients with pulmonary infection of other pathogens in our department from January 1, 2016 to September 30, 2021. Their clinical and imaging data were collected and analyzed. The participants were randomly assigned into training set (n=264) and verification set (n=113) at a ratio of 7 ∶3. Univariate and multivariate analyses were conducted through binary logistic regression, and a nomograph prediction model was constructed using R software package. The validity and application value of the model were evaluated through receiver operating characteristic (ROC) curve, area under ROC curve (AUC), calibration curve, and decision curve (DCA). Results Multivariate logistic regression analysis showed that anemia (OR=6.41, 95%CI: 1.62~25.43), diabetes (OR=6.77, 95%CI: 1.20~38.11), COPD (OR=12.82, 95%CI: 2.71~60.55), malnutrition after admission (OR=8.78, 95%CI: 2.17~35.55), mask/nasal oxygen inhalation (OR=3.53, 95%CI: 0.911~13.66), cholinesterase (OR=7.47, 95%CI: 1.66~33.57), right lung lower lobe (OR=11.17, 95%CI: 2.07~60.16), left lung lower lobe (OR=16.25, 95%CI: 3.31~79.71), ground glass opacity (OR=19.22, 95%CI: 4.17~88.48), air bronchogram (OR=6.44, 95%CI: 1.27~32.68), and bronchiectasis (OR=11.58, 95%CI: 1.50~89.25) were related factors of pulmonary fungal infection. The nomogram was constructed from the above 11 factors (R2=0.830, C-index=0.97, 95%CI: 0.96~0.99). The AUC of the training and verification sets were 0.972 (95%CI: 0.955~0.988) and 0.945 (95%CI: 0.905~0.985), respectively, indicating a high efficiency of differentiation. The calibration curves of the training set and the verification set were distributed along the 45° line basically, and the DCA curves showed that there was a net benefit when the threshold probability was 10%~90%. Conclusion Based on 11 parameters, anemia, COPD, diabetes, malnutrition after admission, mask/nasal oxygen inhalation, cholinesterase, right lung lower lobe, left lung lower lobe, ground glass opacity, air bronchogram sign and bronchiectasis, our prediction model of invasive pulmonary fungal infection is of high accuracy in clinical practice.