Characteristics of Mucormycosis in Hematological Patients and a Death Prediction Model

Abstract

Mucormycosis is an angioinvasive fungal infection, associated with high mortality. The aim of our study was to explore the high-risk factors and predict the death of hematological disease complicated with mucormycosis. We retrospectively analyzed clinical data of 31 patients with hematological disease complicated with mucormycosis, adopted random forest to establish the death prediction model, and validated the model in another 15 patients. The median age of the 31 cases was 46 (28-51) years, male to female ratio 1.38:1, and 90-day mortality rate 54.8%. The most common underlying disease was acute myeloid leukemia (58.1%). The main clinical symptoms were fever (100%), cough (87.1%), sputum (80.6%), chest pain (61.3%), and hemoptysis (19.4%). Reversed halo sign (83.9%) was the most common computed tomography sign. A total of 48.4% of patients also had aspergillus or bacterial infections. Discriminative models were constructed by random forest with 17 non-survivors and 14 survivors. Procalcitonin, the duration of intravenous administration of amphotericin B or amphotericin B liposomes, and neutropenia at death or 90 days of survival were the leading risk factors for poor prognosis, with area under the curve of 0.975 (95% CI 0.934-1). We chose 0.6775 as death prediction threshold (with 82.3% sensitivity and 100% specificity) and validated the model successfully in another 15 patients. Chest pain and reversed halo sign are specific clinical and image signs of hematological disease complicated with mucormycosis. Neutropenia, elevated procalcitonin, and insufficient use time of amphotericin B or amphotericin B liposomes are risk factors for death.

Keywords: death prediction model; hematological diseases; high-risk factors; mucormycosis; random forest.

FIGURE 1

Flow diagram of included patients.

FIGURE 2

Survivorship curves of 46 patients in groups treated with amphotericin B (AMB) or amphotericin B liposome (L-AMB) or not. Survival is significantly lower in patients treated without AMB or L-AMB (P = 0.014).

FIGURE 3

Performance of selected features. X axis is number of features used in random forest leave-one-out cross validation models, y axis is area under curve (AUC). For each random forest model, the prediction process was repeated ten times.

FIGURE 4

Receiver operating characteristic (ROC) curve of leave-one-out cross validation random forest models by incorporating procalcitonin, medication time and neutropenia in training cohort. (Improved group is 14, death group is 17).

FIGURE 5

Probability of death (POD) predicted by using random forest models in training cohort and validation cohort. POD 0.6775 was set as threshold.

FIGURE 6

Dynamic probability of death predicted by using random forest models in training cohort. Each panel indicated one patient. X axis is time points during the whole medication process, y axis is probability of death.