JMIR Publications

ABSTRACT

Background:

Early-stage pulmonary sarcoidosis with serum angiotensin-converting enzyme (sACE) activities within the normal range is difficult to be distinguished from other pulmonary diseases with similar radiographic characteristics.

Objective:

This study aimed to develop and validate machine learning (ML) diagnostic models for differential diagnosis of sarcoidosis using indices from routine peripheral blood tests.

Methods:

From March 2014 and March 2021, 1,142 cases with atypical pulmonary imaging manifestations at the initial hospitalization were enrolled. After applying stringent inclusion and exclusion criteria, a total of 320 cases were utilized to develop two differential diagnostic models using a sex-stratified approach. Feature selection was performed combing Least Absolute Shrinkage and Selection Operator (LASSO) regression with clinical relevance. Six ML algorithms were employed to construct the predictive models, and their diagnostic performance was compared using evaluation metrics, including the area under the receiver-operating-characteristic curve (AUC) and calibration curve. In addition, the clinical decision curve (DCA) was used to evaluate the clinical utility of the models.

Results:

Two models constructed with Support Vector Machine (SVM), for male and female patients with six features each, achieved validation AUCs of 0.888 (male) and 0.825 (female). Implementing further stratification by lymphadenopathy characteristics improved the female model's performance to an AUC of 0.863.

Conclusions:

Based on stratified strategies, the models developed using routine laboratory data and machine learning can effectively and rapidly differentiate early-stage pulmonary sarcoidosis with normal sACE levels..