The prognostic and clinical significance of hemogram parameters: a single-center retrospective analysis

Study population/subjects

The study was run at SBU Izmir Tepecik Training and Research Hospital, a tertiary center, and its affiliated hospital. Ethics approval was obtained (No. 2025/02-23; 10 Mar 2025). We included patients who had their first CBC between 1 Jan and 31 Dec 2024 at the main hospital, while identically filtered data from the affiliated hospital – providing outpatient care and palliative care, obstetric, and gynecologic-oncology units – formed the external test set.

Inclusion criteria

1. Patients who underwent a CBC test.

2. Patients with clearly documented hospitalization status (admitted or not).

3. Patients with at least one complete and valid set of CBC parameters.

4. Patients admitted to the intensive care unit, inpatient wards, or managed on an outpatient basis.

Exclusion criteria

1. Patients with incomplete or erroneous laboratory data.

2. Patients with known chronic hematological diseases.

3. Patients receiving chemotherapy or immunosuppressive therapy.

4. Patients with acute trauma or those who underwent major surgery within the past 3 months.

5. Pregnant or postpartum patients.

6. Patients with chronic kidney disease (≥stage 4) or those undergoing hemodialysis.

7. Patients with congenital or acquired immunodeficiency syndromes.

8. Patients with duplicate CBC test entries during the study period.

Whole blood was drawn into K2-EDTA tubes by ward nurses or phlebotomy nurses and analyzed within 30 min on a Sysmex XN-1000 (Kobe, Japan). All CBC reagents, calibrators, and internal quality control (QC) materials were certified, manufacturer-supplied products.

Study design

The study commenced following the approval of the institutional ethics committee. All patient identifiers were anonymized to ensure confidentiality. A dataset comprising information on treatment type, age, sex, and hemogram results – including white blood cell (WBC), neutrophil (NEU), lymphocyte (LYM), monocyte (MONO), eosinophil (EOS), basophil (BASO), platelet count (PLT), hemoglobin (HGB), red cell distribution width (RDW), mean corpuscular volume (MCV), mean platelet volume (MPV), platelet distribution width (PDW), hematocrit (HCT), as well as calculated ratios such as NLR, PLR and LMR was compiled from 565,877 patients and exported to Microsoft Excel 2021 (USA) for analysis.

HGB and HCT more accurately represent oxygen-transport capacity than red-blood-cell count (RBC), which also lacks prognostic value for acute admission. Plateletcrit (PCT) is a calculated, not measured, parameter (PLT × MPV/10,000). Mean corpuscular hemoglobin (MCH) and its concentration (MCHC) are mainly anemia markers and do not indicate acute severity or systemic inflammation. Therefore, RBC, PCT, MCH, and MCHC were excluded.

After applying the exclusion criteria, the final dataset consisted of 430,561 patients. This dataset was subsequently imported into Python (version 3.11, USA) for machine learning-based analysis. Non-numeric categorical variables were recoded into numerical values as follows: male=0, female=1; care setting: outpatient 0, surgical ward 1, internal ward 2, ICU 3. Outpatients exceeded inpatients ten-to-one, so 10,000 cases per group were randomly sampled. After identical filtering, the affiliated hospital supplied 6,132 records for external testing.

After cleaning, data were stratified by the four-class outcome and randomly split 75:25 into training and internal-test sets, preserving class proportions. Figure 1 – an STARD flow diagram – shows screening, exclusions, final sample sizes, and allocation to training, internal, and external test sets.

Figure 1:

The standards for reporting diagnostic accuracy (STARD) diagram. Cohort selection workflow. A step-wise exclusion of 565,877 candidate samples produces a balanced core dataset of 40,000 cases. The data are then split into training, internal test and external test cohorts (3:1:≈0.2). Sub-category counts (ICU, internal medicine, surgical, outpatients) are annotated in each block.

Data preprocessing and training of machine learning algorithms

The cleaned dataset was imported into the Python programming environment (v3.11) using PyCharm IDE. Model development and evaluation were performed with widely used libraries such as TensorFlow and scikit-learn, enabling efficient data preprocessing, training, and optimization [10].

We tested five classifiers – multinomial logistic regression (MN-LR), random forest (RF), gradient boosting (GBM), support-vector machines (SVM), and a multilayer perceptron neural network (MLP-ANN). All are proven for structured clinical data; ensemble (RF, GBM) and neural-network (MLP-ANN) models capture the nonlinear, high-dimensional patterns typical of biomedical datasets.

All procedures were conducted within the Python 3.11 programming environment. The libraries utilized are categorized as follows:

1. Data Processing and Analysis: pandas (v1.5), numpy (v1.23).

2. Machine Learning Model Development: scikit-learn (v1.2), xgboost (v1.6), lightgbm (v3.3), tensorflow (v2.10), keras (v2.10).

3. Model Evaluation and Visualization: matplotlib (v3.6), seaborn (v0.12), scipy. stats (v1.9), sklearn. metrics (v1.2), mlxtend (v0.21), shap (v0.47).

Model hyperparameters were optimized via grid-search cross-validation, and performance was subsequently assessed on an independent, stratified test set. Residual class imbalance was mitigated by applying inverse-frequency class weights during training, thereby limiting majority-class bias and ensuring balanced discrimination across outcome categories.

Following model training, performance evaluation was conducted using the designated test dataset.

Performance evaluation

The modeling process underwent comprehensive evaluation, including hyperparameter tuning and model selection through internal cross-validation. Model performance was assessed using multiple evaluation metrics. The following criteria were used for classification:

1. Classification Performance Metrics

   1. Area Under the Receiver Operating Characteristic Curve (AUC-ROC)

   2. Area Under the Precision-Recall Curve (AUC-PR)

   3. Confusion Matrix Analysis

   4. Sensitivity, Specificity, Positive Predictive Value (PPV), and Negative Predictive Value (NPV), F1 Score

2. Model Interpretability Metrics

   1. Feature Importance Analysis

   2. Two-Dimensional Decision Boundary Visualization

   3. SHAP (Shapley Additive Explanations) Graphs: SHAP values were employed to enhance model transparency and enable instance-level interpretation of feature contributions

3. Validation Results of the Predictive Models (with external test set): were analyzed to ensure comprehensive assessment. This structured and multifaceted evaluation approach provides a robust framework for predicting treatment modality outcomes based on laboratory-derived data.

Statistical analysis

Statistical analyses were performed using Python (v3.11). Continuous variables were summarized as means and standard deviations, while categorical variables were expressed as counts and percentages. Differences between the development and external test sets were assessed using independent samples t-test for age and Pearson Chi-square test for sex distribution.

To compare hemogram parameters across treatment groups (outpatient, surgical ward, internal medicine ward, ICU), the Kruskal–Wallis test was applied for continuous variables due to non-normal distribution. For categorical comparisons, the Pearson Chi-square test was used. A two-tailed p-value of <0.05 was considered statistically significant.

Dataset description and data pre-processing

The study included a total of 40,000 data points derived from hospital records. The dataset consisted of hemogram parameters in conjunction with demographic characteristics of the patients. In addition, a preprocessed external test comprising 6,132 records was appended to the main dataset. Baseline characteristics of the study population, including main dataset (train and internal test set) and external test set, are summarized in Table 1, while the descriptive statistics of the hemogram and related variables are presented in Table 2. When assessing the statistical significance of differences in variable means across treatment types, the following comparisons were found to be non-significant: for lymphocyte count in treatment types 0 vs. 2 and 0 vs. 3 within the training set; and for lymphocyte (0 vs. 3), platelet (0 vs. 2), and mean platelet volume (0 vs. 1 and 0 vs. 2) within the test set. All other pairwise comparisons between treatment types demonstrated statistically significant differences in mean values (p<0.05). The main dataset (mean age 48.7 years) contained slightly older women than men (50.3 vs. 47.3 year) and more females overall (52.6 %). Male density peaked at 66–80 years and fell >80 years. External-test patients were significantly older – as they came from palliative and gynecologic-oncology wards – and included a higher proportion of women (p<0.05). A detailed breakdown of age distribution by gender is illustrated in Figure SM1 (Supplementary Material 1).

For the variables incorporated into the predictive models, Pearson correlation analysis was applied to assess linear relationships, whereas Spearman rank correlation was used to evaluate monotonic associations. The correlation matrices and corresponding visualizations for these analyses are presented in Figure SM2 (Supplementary Material 1).

The performance of MN-LR, RF, GBM, MLP-ANN, and SVM models was comparatively evaluated based on their predictive capabilities, classification metrics, and interpretability. Classification metrics such as F1 score, sensitivity, specificity, and AUC-ROC were used to assess the models’ ability to discriminate between classes.

Each algorithm offers distinct advantages and limitations depending on the nature of the dataset and the classification problem. MN-LR is valued for its simplicity and computational efficiency; RF and GBM provide robust predictive performance and are less prone to overfitting; MLP-ANN is particularly effective for large and complex datasets; and SVM excels in smaller, well-separated datasets. Model selection was guided by the structural characteristics of the data and the specific classification objectives. All models were systematically analyzed to reflect the diverse methodological strengths of each algorithm.

Comparison of classification performance metrics

In this section, MN-LR, RF, GBM, MLP-ANN, and SVM were compared on sensitivity, specificity, PPV, NPV, and F1-score. Metrics that satisfied normality and equal-variance criteria (e.g., F1-score) were analyzed with one-way ANOVA, while non-normal metrics (e.g., sensitivity) used the Kruskal–Wallis test to detect performance differences among models.

All models outperformed MN-LR in nearly every treatment group. Performance was strongest for outpatients: RF and SVM best distinguished them from inpatients. GBM and SVM excelled for ICU cases, whereas accuracy for internal and surgical ward patients remained moderate across models.

Although MN-LR showed the lowest mean F1 (0.60 ± 0.14), its score did not differ significantly from RF (0.64 ± 0.14), GBM (0.68 ± 0.13), MLP-ANN (0.67 ± 0.13), or SVM (0.68 ± 0.12); all t-test p-values exceeded 0.47. Likewise, GBM, MLP-ANN, and SVM displayed no pairwise differences (p>0.92). AUC-ROC and AUC-PR curves nevertheless highlighted GBM and MLP-ANN as the most accurate models across treatment groups. Detailed metrics appear in Table 3 and Figure 2, with performance curves for the other models presented in Supplementary Material 1.

Figure 2:

AUC-ROC and AUC-PR plots of the methods. Discriminative performance via ROC and PR curves. Top two rows: class-specific ROC and PR curves for GBM and MLP-ANN (colors denote treatment classes). Bottom row: macro-averaged ROC (left) and PR (right) curves comparing all models; the grey dashed line indicates random performance. 0, outpatients; 1, surgical services; 2, internal services; 3, intensive care units; NM-LR, multinomial logistic regression; RF, random forest; SVM, support vector machine; AUC-ROC, area under the receiver operating characteristic curve; AUC-PR, area under the precision-recall curve.

Based on the overall AUC-ROC and AUC-PR values, the models were ranked from highest to lowest performance as follows: GBM, MLP-ANN, SVM, MN-LR, and RF. The GBM model, which achieved the highest AUC values, yielded 26,748 true negatives (TN), 6,748 true positives (TP), 3,252 false negatives (FN), and 3,252 false positives (FP). In comparison, the MLP-ANN model correctly classified 26,588 TN and 6,588 TP, while misclassifying 3,412 FN and 3,412 FP. These results indicate that although both models performed well, GBM had slightly better overall classification accuracy due to fewer false predictions.

Decision-boundary plots highlighted contrasting architectures: GBM formed sharp, piecewise-linear splits characteristic of tree models ideal for tabular data, whereas MLP-ANN produced smoother, continuous transitions that capture complex nonlinear patterns – explaining its aptitude for high-dimensional feature spaces.

The confusion matrices and decision boundary plots for both GBM and MLP-ANN models are presented in Figure 3. Corresponding visualizations for the other models are provided in Supplementary Material 2.

Figure 3:

Confusion matrix and decision boundary plots. Confusion matrices and decision boundaries for GBM and MLP-ANN. Left column (confusion matrix): True-vs-predicted counts across four treatment classes. Deep diagonal blues indicate correct classifications; off-diagonals reveal residual confusion. GBM achieves its best recall on class 0, whereas the ANN favours class 1. Right column (decision boundary): test-set points projected onto the first two PCA components; background shading depicts the multi-class decision surface. The GBM exhibits block-like partitions, while the MLP-ANN displays smoother, radial regions – consistent with tree-based vs. neural-network decision mechanisms. GBM, gradient boosting machine; MLP-ANN, multilayer perceptron – artificial neural network.

Validation results of the models (with external test set)

In accordance with the International Federation of Clinical Chemistry and Laboratory Medicine (IFCC) recommendations, external validation was conducted to ensure the generalizability and robustness of our models [11]. External validation centered on the top performers, GBM and MLP-ANN. Both retained strong results across all care categories (Table 3). GBM yielded F1 scores of 0.92, 0.70, 0.69, 0.77 for types 0–3 (macro-F1 = 0.77) and AUC-ROC/AUC-PR ranges of 0.79–0.95/0.55–0.86, confirming reliable discrimination, particularly for outpatients and ICU admissions.

Similarly, MLP-ANN yielded F1 scores of 0.90, 0.69, 0.67, and 0.74 (macro-F1 = 0.75), with AUC-ROC 0.78–0.93 and AUC-PR 0.53–0.83. GBM remained marginally superior, especially in ICU classification.

Both models generalized well to the external cohort, yet prospective multi-center studies are required before routine clinical use. Confusion matrices for GBM and MLP-ANN in Supplementary Material 2 confirm balanced performance across outcome categories.

Comparison of interpretability performance metrics

In both of the top-performing models – GBM and MLP-ANN – EOS count emerged as the most influential predictor, as illustrated in the feature importance rankings. For the GBM model, EOS was followed by BASO, age, and HCT, whereas in the MLP-ANN model, HCT, HGB, and WBC count were the next most important variables. Notably, the widely studied hemogram-derived ratios NLR, PLR and LMR – were ranked only as moderate to low contributors in both models.

Although eosinophil count is not typically regarded as a key clinical parameter for determining hospitalization needs, its high relative importance in both models suggests a statistically significant – though potentially non-causal – association with treatment classification outcomes. This highlights the ability of machine learning models to uncover patterns that may not be immediately apparent in traditional clinical assessments.

Figure 4 presents the feature importance plots and SHAP graphs, which collectively demonstrate both the magnitude and direction of each variable’s influence on individual predictions. The contrast between the GBM and ANN models in their handling of feature interactions reflects their differing algorithmic structures – GBM’s tree-based logic vs. ANN’s capacity for capturing nonlinear relationships.

Figure 4:

Feature importance and SHAP plots global and local feature importance for GBM vs. MLP-ANN. The upper bar plot contrasts global importance scores (aggregate SHAP values; GBM blue, ANN orange). Bottom violin/“bee-swarm” summaries depict local SHAP distributions and sign (blue=low, pink=high) for each model. Features are ordered by model-specific importance. SHAP, shapley additive explanations; GBM, gradient boosting machine; MLP-ANN, multilayer perceptron – artificial neural network; WBC, white blood cell; NEU, neutrophil count; LYM, lymphocyte count; MONO, monocyte count; EOS, eosinophil count; BASO, basophil count; PLT, platelet count; HGB, hemoglobin concentration; RDW, red cell distribution width; MCV, mean corpuscular volume; MPV, mean platelet volume; PDW, platelet distribution width; HCT, hematocrit; NLR, neutrophil-lymphocyte ratio; PLR, platelet-lymphocyte ratio; LMR, lymphocyte-monocyte ratio.

Furthermore, the Matthews Correlation Coefficient (MCC) values were calculated as 0.57 for GBM and 0.55 for MLP-ANN, indicating a moderate level of classification performance for both models.

Strengths and limitations

A key strength of this study lies in its large, heterogeneous patient population, drawn from a tertiary-level training and research hospital situated in a densely populated urban center with high patient turnover. This robust sample enhances the statistical power and relevance of the findings. Furthermore, the comparative evaluation of multiple machine learning algorithms provides a multifaceted understanding of predictive performance across diverse modeling approaches.

Nevertheless, the retrospective and single-center design may limit the generalizability of the results. To enhance external validity and promote clinical translation, future studies should adopt prospective, multicenter designs. Comparative assessments that integrate machine learning predictions with expert clinical judgment across varied clinical scenarios may also improve real-world applicability. Additionally, combining hemogram-derived ratios with other biochemical and clinical variables could yield more comprehensive and accurate prognostic models.