Impact of serum calcium levels on the occurrence of sepsis and prognosis in hospitalized patients with concomitant psoriasis: a retrospective study based on the MIMIC-IV database

Abstract

Objective: This study aims to investigate the relationship between serum calcium levels during hospitalization and the incidence and prognosis of sepsis in hospitalized patients with psoriasis.

Methods: A retrospective analysis of patients with concomitant psoriasis admitted for the first time was conducted, utilizing the Medical Information Mart for Intensive Care database. Machine learning techniques, along with logistic regression, Cox regression, group-based trajectory modeling (GBTM), and mediation analysis, were employed to assess the influence of serum calcium levels and other clinical indicators on the occurrence of sepsis and all-cause mortality.

Results: Serum calcium exhibits a significant inverse correlation with the occurrence of sepsis [odds ratio (OR) =0.351, 95% CI: 0.265-0.463, P<0.001]. Furthermore, serum calcium levels exhibited a negative correlation with 90-day all-cause mortality [hazard ratio (HR)=0.594, 95% CI: 0.422-0.835, P=0.003] and a similar negative correlation with 365-day mortality risk (HR=0.642, 95% CI: 0.502-0.821, P<0.001). Platelet counts mediated the relationship between serum calcium and both 90-day and 365-day all-cause mortality, accounting for 24.6% and 22.0% of the mediation effect, respectively. Additionally, three distinct trajectory patterns based on serum calcium levels were identified, with the low calcium trajectory group exhibiting a higher risk of sepsis (OR=2.400, 95% CI: 1.163-5.068, P<0.001).

Conclusion: Serum calcium levels serve as a significant predictive factor for the occurrence and prognosis of sepsis in hospitalized patients with psoriasis. Continuous monitoring of serum calcium levels and timely correction of hypocalcemia may contribute positively to improving patient outcomes.

Keywords: machine learning; mediation analysis; psoriasis; sepsis; serum calcium; trajectory modeling.

Figure 1

Flowchart of patient inclusion in the study.

Figure 2

(A) Lasso regression for feature selection, six influencing factors were screened out, with a λ value (1se) of 0.035. The characteristic variables include platelet count, white blood cell count (WBC), serum calcium, urea nitrogen (UN), anion gap, and the use of glucocorticoids. (B) The constructed nomogram model based on these selected features. From the values of the six predictive factors in the chart, trace vertically upwards to intersect with the ‘Points’ line above. The scores for each predictive factor are summed to obtain a total point, and then refer to the “Total Points” axis to read the corresponding scale on the “probability of sepsis” axis below, so as to predict the risk of sepsis occurrence.

Figure 3

Evaluation of the predictive ability of the nomogram model (A) ROC curve of the training set, the AUC of nomogram was 0.813 (95%CI 0.769-0.857). (B) Calibration curve of the training set, “Ideal” represents the ideal reference line of the nomogram. “Apparent” represents the actual performance of the nomogram, while “Bias - corrected” represents the performance of the nomogram after correcting for bias. (C) Decision curve analysis (DCA) curve of the training set. (D) ROC curve of the validation set, the AUC of nomogram was 0.763 (95%CI 0.703-0.823). (E) Calibration curve of the validation set. (F) DCA curve of the validation set.

Figure 4

(A) Bee swarm plot of SHAP values calculated using the random forest algorithm. Each point represents an observation. The x - coordinate of the point is the SHAP value, and the color gradient of the point corresponds to the value of the independent variable. The arrangement of independent variables on the y - axis is consistent with the ranking of SHAP variable importance, that is, from top to bottom, the importance of independent variables decreases. (B) The Restricted Cubic Spline (RCS) curve analysis indicates that there is a significant non - linear relationship (P <0.001) between serum calcium and the occurrence of sepsis. UN, Urea nitrogen.

Figure 5

Comparison of predictive capabilities of machine learning algorithms (A) ROC curve, the logistic regression algorithm demonstrated the highest AUC value (0.742) in the comparison. (B) PR curve, the horizontal axis represents the recall rate, and the vertical axis represents the precision rate. The random forest (AUC = 0.441) and logistic regression (AUC = 0.428) had relatively better performance. Their curves were closer to the upper - right corner, indicating better performance in balancing precision and recall.

Figure 6

Analysis of the mediating role of platelets in the relationship between serum calcium and all-cause mortality (A) Analysis of 90-day all-cause mortality; (B) Analysis of 365-day all-cause mortality. Adjusted for age and gender.

Figure 7

Patterns of serum calcium changes after hospital admission. Trajectory 1, slow increase trend; Trajectory 2, decreasing -increasing trend; Trajectory 3, decreasing trend.