Both TTP and Immune Thrombocytopenia (ITP) present with thrombocytopenia in clinical practice, which may be related to inflammatory factors. In order to improve the specificity of TTP diagnosis, highlight the characteristics of TTP related inflammatory factors, and improve the clinical practicality of the prediction model, we selected TTP and ITP patients as control subjects for the study. This retrospective investigation was performed in the Second Affiliated Hospital of Guangzhou Medical University in Guangdong Province, China. A total of 254 consecutive patients with thrombocytopenia admitted between May 22, 2019, and May 22, 2024, were included. This study adhered to the principles of the Declaration of Helsinki and received approval from the Clinical Research and Application Ethics Committee of the Second Affiliated Hospital of Guangzhou Medical University. The inclusion criteria consisted of participants aged ≥18 years and<95 years, who underwent regular blood tests within 24 h of admission; and patients objectively diagnosed with TTP or ITP upon initial admission, possessing comprehensive clinical data. The exclusion criteria encompassed pregnant women, individuals with solid or haematologic malignancies, those with other thrombotic diseases, and patients using immunosuppressants (18). According to the “Chinese Guidelines for the Diagnosis and Treatment of Primary Immune Thrombocytopenia in Adults (2020 Edition).” The diagnosis of ITP is mainly based on clinical exclusion, requiring a decrease in platelet count, generally no splenomegaly, and bone marrow cell morphology characterized by increased or normal megakaryocytes with maturation disorders etc. (19). According to the “Chinese Guidelines for the Diagnosis and Treatment of Thrombotic Thrombocytopenic Purpura (2022 Edition).” The diagnosis of TTP, after excluding other thrombotic microangiopathy, includes: significantly decreased platelet count, fragmented red blood cells and increased reticulocyte ratio in peripheral blood smear; increased blood bilirubin, significantly increased lactate dehydrogenase (LDH), increased blood urea nitrogen and creatinine, and detection of ADAMTS13 activity and inhibitors etc. (20). A total of 196 eligible patients were categorized into the TTP group and the ITP group according to the final diagnosis. The patient screening process was illustrated in Figure 1.

Diagnostic workup included tests for antiplatelet membrane glycoprotein autoantibodies, antinuclear autoantibody group, antiphospholipid antibodies, hepatitis virus serology, thyroid function tests, serum immunoglobulin levels, serum thrombopoietin levels, and genetic testing (20). All diagnoses were centrally assessed by a minimum of one chief physician from the haematology department and one laboratory technician, both of whom were uninformed of other outcomes. Demographic information, fundamental anthropometric data, medical history, clinical characteristics and laboratory test results (mainly blood routine test results), past medical history including that of hypertension, diabetes and other disorders of the patients were collected within 24 h after admission. SII, PLR, PPN, and body mass index (BMI) were calculated according to the following calculation formulas: SII = (neutrophil count × platelet count) / lymphocyte count; PLR = platelet count/lymphocyte count; PPN = neutrophil count × platelet count; BMI = weight (kg) / height2 (m2).

Continuous variables were presented as mean ± standard deviation (SD) or median (interquartile range, IQR), with normality assessed using the Kolmogorov Smirnov test. Between group comparisons were conducted using t-tests or Mann Whitney U tests, while categorical variables were analyzed with chisquare tests. A two-tailed p-value < 0.05 was considered statistically significant. Univariate logistic regression analysis was employed to identify risk factors associated with TTP. All statistical analyses were performed using SPSS software version 21.0.

Feature selection was based on both clinical relevance and statistical significance, with a particular focus on inflammatory markers (SII, PLR, and PPN). These markers have been implicated in thrombosis and immune dysregulation, making them physiologically relevant for distinguishing TTP from ITP. To our knowledge, this was the first study to integrate these inflammatory indices into a machine learning model for TTP diagnosis.

To develop and validate a diagnostic model, data from 196 patients were randomly divided into a training set (90%) and a test set (10%). To mitigate overfitting from limited sample size, we implemented 10-fold cross-validation (stratified by class) with fixed random seed (500). Eight machine learning models were implemented, including Naïve Bayes, XGBoost, Support Vector Machine (SVM), Random Forest, K-Nearest Neighbors (KNN), Gradient Boost Trees, Decision Tree, and Logistic Regression (21, 22). All models were applied to the binary classification task of differentiating TTP from ITP. Model performance was averaged across folds using receiver operating characteristic curve (AUC), recall, F1 score, accuracy, and confusion matrix. All machine learning models were implemented in Python (version 3.11.10) using PyCharm as the development environment. Confusion matrices were generated to compare predicted and actual labels, with grid values representing true negatives (TN), false positives (FP), false negatives (FN), and true positives (TP), thereby enabling direct assessment of each model’s sensitivity and specificity.

Given the urgent need for early TTP diagnosis, AUC was prioritized to minimize false negatives and reduce the risk of missed diagnosis. Model risk factors were visualized using nomograms, model reliability was assessed using calibration curves, and the net clinical benefit of applying the model to decision making was determined using decision curve analysis (DCA) (23).

After comprehensive evaluation, we selected logistic regression as the final model because it achieved the highest balance between accuracy (0.78), interpretability, and clinical feasibility. The feature contribution was visualized using SHAP values. Unlike black-box models such as XGBoost or SVM, logistic regression can directly calculate risk scores, making it more suitable for real world clinical practice.

This study consisted of 196 patients, including 118 females (60.20%) and 78 males (39.80%), with an average age of 47.60 ± 16.74 years, and a BMI of 23.20 ± 3.44. Among the participants, 28 had a history of smoking (14.29%), 14 had a history of alcohol consumption (7.14%), 37 had diabetes (18.88%), 41 had hypertension (20.92%), and 8 had coronary heart disease (4.08%). The analysis of essential laboratory test results (platelet count, lymphocyte count, neutrophil count) indicated that SII ranged from 192.51 ± 221.00, PPN ranged from 224.86 ± 209.99, and PLR ranged from 30.13 ± 32.96. Table 1 enumerates specific demographic and clinical factors.

The patient’s baseline characteristics were shown in Table 2. Among them, 54 were patients with TTP and 142 were patients with ITP. Parameters such as body weight (p = 0.023), BMI (p = 0.018), neutrophil count (p = 0.043), platelet count (p < 0.001), SII (p < 0.001), PPN (p < 0.001), and PLR (p < 0.001) exhibited significant differences between the ITP group and the TTP group (p < 0.05). Although the mean BMI was higher in the TTP group, there was substantial overlap in the distributions between groups, suggesting that BMI alone may not reliably distinguish TTP from ITP and should be considered in combination with inflammatory indices.

Given that BMI calculation incorporates weight and height, and combined inflammatory index calculation encompasses neutrophil, platelet, and lymphocyte count parameters, in order to avoid duplication, the subsequent studies excluded weight, height, and individual laboratory indicators to prevent redundancy.

Table 3 illustrates the compositional differences of variables between the two groups. Significant differences in BMI, SII, PPN, and PLR (p < 0.05) between the groups were determined as risk factors. Among them, BMI was a protective factor (OR = 0.89, 95% CI (0.80, 0.98), p = 0.019), whereas SII, PPN, and PLR were identified as risk factors [OR = 1.01, 95% CI (1.00, 1.01), OR = 1.00, 95% CI (1.00, 1.01), OR = 1.03, 95% CI (1.02, 1.05), respectively].

The data were randomly allocated to the training set and the test set in a 9:1 ratio utilizing Python. The internal validation employed 10-fold cross-validation with a seed number of 500 to guarantee the consistency of all parameters. Eight machine learning models were constructed (Table 4): Naive Bayes, Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Random Forest, K-Nearest Neighbors (KNN), Gradient Boosting Tree, Decision Tree, and Logistic Regression. All models were applied to the binary classification task of differentiating TTP from ITP. Together with Area under curve (AUC), recall, F1 score, accuracy, and confusion matrix were established as performance evaluations. The receiver operator characteristics (ROC) curve and confusion matrix was illustrated (Figures 2, 3) to enhance the intuitiveness and clarity of the model’s prediction ability.

The diagnostic performance of the eight machine learning models was summarized in the confusion matrices shown in Figure 3. Each matrix illustrates the numbers of true negatives (TN), false positives (FP), false negatives (FN), and true positives (TP), providing a visual complement to summary metrics. While several models, such as Random Forest, XGBoost, and Decision Tree, demonstrated relatively high true negative counts, they yielded few true positives, resulting in higher false negatives and thus lower recall and F1 scores—an especially critical limitation in clinical contexts where minimizing missed TTP diagnoses is essential. Logistic regression, by contrast, achieved the most balanced distribution between true positives and true negatives, consistent with its superior performance across multiple evaluation metrics. Specifically, it yielded the highest AUC (0.78; 95% CI: 0.71–0.85), recall (0.46), F1 score (0.55), and accuracy (0.78), outperforming more complex classifiers including SVM, Gradient Boosting, and KNN (Figure 2; Table 4). These results, supported by both numerical metrics and the confusion matrix visualization, highlight logistic regression as the most clinically reliable and interpretable model for early TTP risk prediction. Notably, the discrimination between TTP and ITP was primarily driven by composite inflammatory markers (SII, PPN, and PLR), with logistic regression leveraging these features to achieve the most reliable performance.

Consequently, with the logistic regression model, the fitting curve was generated by 1,000 repeated extractions. A calibration curve (Figure 4) and a Decision Curve Analysis (DCA) curve (Figure 5) were constructed. The results indicated that the calibration curve’s real fit was consistent with the ideal fit, demonstrating that the predicted probability matched the actual probability. The clinical decision curve exhibited a substantial clinical benefit rate, further elucidating the applicability of this approach in clinical settings. So, a nomogram (Figure 6) prediction model was developed to visualize risk factors and assess TTP risk. The four variables were assigned the four scores in the nomogram. The probability could be directly obtained by adding the scores of each predictor, further visualizing the clinical data of high risk TTP patients, boosting diagnostic accuracy.

The Shapley additive explanation (SHAP) and AUC were used to visually display the influence of the selected variables on the diagnosis of TTP. Figure 7 illustrates the classification of the four variable features in the model, plotting points of different colors corresponding to each feature variable. It showed that SII is the main risk factor affecting the diagnosis of TTP, and confirms that SII is the most influential feature for distinguishing TTP from ITP, and its predictive ability exceeds that of PPN and PLR. In addition, ROC calculation (Figure 8) showed that the ROC of SII = 0.777 (95% CI = 0.706, 0.849), which exceeds the ROC of PPN 0.776 (95% CI = 0.701, 0.851). SII was identified as the most influential predictor in our model, consistent with its role in systemic inflammation and platelet activation.