Common laboratory results-based artificial intelligence analysis achieves accurate classification of plasma cell dyscrasias

Patients and methods

The study adhered to the principles of the Declaration of Helsinki and received approval from the Medical Ethics Committee of Zhejiang Provincial People’s Hospital (Approval No. Zhejiang Provincial People’s Hospital Ethics 2024 Other No. 034, Acceptance No. QT2024029). Given the retrospective design, the Ethics Committee granted a waiver for the requirement of individual informed consent.

Patients and data selection

A retrospective analysis was conducted on common and biochemical markers from 609 newly diagnosed plasma cell dyscrasia cases and 579 control cases (comprising infectious diseases, autoimmune disorders, liver diseases, and kidney diseases) treated at Zhejiang Provincial People’s Hospital between January 2018 and February 2024. To further evaluate the model’s clinical applicability and generalizability, data from an additional 30 newly diagnosed plasma cell dyscrasia cases and 34 control cases were included. Diagnoses followed the 2014 International Myeloma Working Group (IMWG) criteria. Thirteen variables—sex, age, hemoglobin (HGB), calcium, serum creatinine, erythrocyte sedimentation rate (ESR), κ light chain, λ light chain, κ/λ ratio, albumin/globulin ratio, albumin, globulin, and β₂-microglobulin—were selected as initial modeling variables, representing key biomarkers associated with the CRAB criteria. These markers have been validated as independent diagnostic indicators for plasma cell dyscrasias (Cowan et al., 2022), aiding in early detection. For subtype classification, additional data on immunofixation electrophoresis and CD markers were utilized. A deep learning model was employed, optimizing classification accuracy and effectively distinguishing between plasma cell dyscrasia subtypes.

Despite the essential role of radiological assessments, such as X-ray, whole-body MRI, and PET-CT, in diagnosing plasma cell dyscrasias, these modalities were deliberately excluded from the analysis. The focus was on developing a model based solely on routine laboratory parameters, which are more accessible and cost-effective, allowing for broader and faster clinical implementation. Incorporating radiological data could introduce variability due to differing resource availability, potentially compromising model performance. The emphasis on routine laboratory markers aimed to preserve simplicity and clinical practicality, though it is acknowledged that the exclusion of radiological data may limit the model’s ability to capture certain disease-specific features.

Methodology for evaluating different variables

Demographic and clinical parameters: Sex and age were documented at the time of diagnosis.

Biochemical marker evaluation: HGB, calcium ions (Ca²⁺), serum creatinine, ESR, κ and λ light chains, albumin, globulin, and β₂-microglobulin were quantified using standard clinical laboratory techniques. Automated analyzers were utilized for HGB and calcium measurements, nephelometry for light chain assessment, and ELISA for β₂-microglobulin detection.

Flow cytometry (MFC) analysis: Multiparametric flow cytometry (MFC) was employed to assess CD markers, including CD45, CD38, and CD138, using an antibody panel with 8–12 fluorochromes. Antigen expression was classified as positive when more than 10% of plasma cells demonstrated expression. Table S1 provides detailed antibody-fluorochrome combinations, and further information on tube combinations is available in Table S2.

Data processing

Based on diagnostic criteria and physicians’ auxiliary judgments, factors relevant to diagnosing plasma cell dyscrasias were identified. Raw data for diagnostic prediction were extracted from the LIS database, but the initial dataset required additional preprocessing before it could be used to train machine learning algorithms.

Missing value handling

A significant portion of missing values, especially in immunoglobulin and immunophenotyping data, was observed among non-plasma cell dyscrasia patients. As immunophenotyping and immunoglobulin tests could influence the model, two corresponding data columns were removed. Missing values were predominantly found in the ESR and β₂-microglobulin features, which were handled using mean imputation—replacing missing values with the feature’s average. For categorical features, mode imputation was applied. In addition to these imputation techniques, attention-based deep learning algorithms were used to encode different features. Missing values were labeled as exceptional and masked to enable distinct feature encoding, preserving more information compared to conventional imputation methods.

Categorical feature processing

Machine learning and deep learning models require numerical input features. Therefore, categorical variables were converted into numerical formats. A direct numerical assignment approach was avoided for immunofixation electrophoresis, as this could introduce misleading intensity information. Instead, one-hot encoding (Zhu, Qiu & Fu, 2024) was applied, breaking immunofixation electrophoresis into five binary columns: G, A, M, κ, and λ. Each column was represented as a binary value, where 0 indicated a negative result and one indicated a positive one, converting immunofixation electrophoresis data into a 5-dimensional binary vector.

Variable selection

To develop a robust classification system and identify key laboratory variables specific to plasma cell dyscrasias, a systematic analysis was conducted, comparing data between plasma cell dyscrasias and non-plasma cell dyscrasias. The Kruskal-Wallis (K-W) test or Wilcoxon rank-sum test, along with multiple logistic regression, were employed to evaluate the differential expression of laboratory variables between the two groups. Forest plots were used for visual comparison. Through final dimensionality reduction, nine critical variables were identified for the diagnostic model: HGB, age, serum creatinine, ESR, κ/λ light chain ratio, κ light chain, λ light chain, globulin, and β₂-microglobulin. All statistical tests were considered significant at p < 0.05 (two-tailed).

Patient characteristics

This study focuses on classifying plasma cell dyscrasias and non-plasma cell dyscrasias to develop predictive diagnostic models. A comparative analysis of clinical features and outcomes between individuals with plasma cell dyscrasias and non-plasma cell dyscrasias in the derived cohort is presented in Table 1. Over the study period, 1,188 patients were divided into training and internal validation sets, while an external group of 64 patients was used for validation.

Diagnostic model development

Given the model’s diagnostic and classification capabilities, gradient boosting decision trees (GBDT) are recognized as a highly effective ensemble learning method (Hastie et al., 2009). As a commonly used algorithm in ensemble learning, GBDT was implemented to predict whether patients are affected by plasma cell disorders. GBDT is a boosting-based technique designed for tabular prediction tasks, and it operates as an additive model consisting of multiple CART decision regression trees. The model works by iteratively constructing decision trees, where k represents the number of trees, and $\mathrm{T}\left(\mathrm{x},{\mathrm{\theta }}_{\mathrm{i}}\right)$ denotes the output of the i-th regression tree. In binary classification, ${\mathrm{f}}_{\mathrm{k}}$ is selected based on the mean squared error (MSE) between ${\mathrm{f}}_{\mathrm{k}}$ and the labels as the loss function L. The negative gradient of the total loss J is calculated to fit each successive regression tree ${\mathrm{T}}_{\mathrm{k}}$(x), and these trees are ultimately combined to generate the final prediction.

(1) $$f_k=\sum _{i=1}^{k}T\left(x,{\theta }_i\right)$$

(2) $$J=\sum _{n=1}^{N}L\left(y_n,f_k\left(x_n\right)\right)=\sum _{n=1}^{N}L\left(y_n,f_{k-1}\left(x_n\right)+T_k\left(x\right)\right)$$

The maximum number of regression trees was set to 100, with a minimum leaf size of 2 and a learning rate of 0.1. Preprocessing steps included data imputation, cleaning, and encoding of categorical features. Parameter tuning was performed using 5-fold cross-validation. For external validation, 30 newly diagnosed plasma cell dyscrasias cases and 34 non-plasma cell dyscrasias cases were selected. The GBDT classifier, trained on the original dataset, demonstrated strong generalization, proving both clinically practical and applicable. The full training pipeline is illustrated in Fig. 1.

Performance evaluations were also conducted by comparing the GBDT model with support vector machines (SVM) (Rabbani et al., 2022), deep neural networks (DNN) (Li et al., 2020), and decision trees (DT). The SVM model was designed to identify the hyperplane that maximizes the margin in feature space. A penalty coefficient of one was selected, and the Gaussian kernel function was applied, with the kernel coefficient set to the reciprocal of the feature dimension.

For the DNN model, a deep learning architecture incorporating attention mechanisms was utilized. Details on the structure and benefits of this model will be discussed in subsequent sections. The predictive models were constructed and fine-tuned using 5-fold cross-validation on internal data, followed by external validation to further assess model performance.

Establishment of a subtype classification model

Deep learning models excel at learning data representations across multiple levels of abstraction through several processing layers (Zhu, Qiu & Fu, 2024). Leveraging the backpropagation algorithm, these models process large datasets to identify complex patterns and optimize internal parameters. These parameters are essential for calculating representations in each layer, building on those from the previous layer. In the context of GBDT, each iteration constructs a new decision tree, ${\mathrm{T}}_{\mathrm{k}}$(x), by fitting the model to residuals, effectively reducing bias and improving the model’s generalization capabilities. GBDT remains the state-of-the-art (SOTA) approach for tabular prediction tasks due to its aptitude for handling tabular data with non-smooth decision boundaries (Hastie et al., 2009). However, its performance depends heavily on selecting parameters like the number and depth of decision trees, and it is less suited for high-dimensional sparse datasets.

Self-Attention and Intersample Attention Transformer (SAINT) represents an advanced deep learning approach specifically designed for tabular data, outperforming traditional gradient-boosting methods by integrating sophisticated attention mechanisms and contrastive self-supervised pre-training techniques (Somepalli et al., 2021). SAINT leverages sample-level attention mechanisms to deepen the model’s understanding of feature interactions, enhancing its learning efficiency in the presence of limited labeled data, which leads to significantly improved classification accuracy.

In this study, the task involves handling various data features, such as continuous variables (e.g., creatinine and calcium ions) and categorical variables (e.g., CD markers and immunofixation electrophoresis). Traditional machine learning algorithms typically convert these features into numerical values via one-hot encoding or direct mapping, often failing to capture the latent information within categorical variables. To address this, word embedding techniques commonly used in natural language processing are employed to project categorical features into high-dimensional vector spaces within the embedding layer. For continuous features, mapping into a distinct high-dimensional space is accomplished through two fully connected layers and ReLU activation functions.

At the feature extraction stage, the model introduces attention modules, normalization, and feature mapping layers. The attention mechanism calculates the similarity between the Query and Key across different features, predicting the Query’s value based on the Key’s values. This approach improves the model’s ability to identify correlations between features of varying dimensions. Layer normalization is used to normalize batch outputs, ensuring equal access to all Keys. Additionally, skip connections are incorporated to mitigate the vanishing gradient problem, thereby enhancing both model stability and performance.

(3) $$Attention\left(Q,K,V\right)=Softmax\left({\displaystyle \frac{Q{K}^T}{\sqrt{d_k}}}\right)V.$$

At diagnosis, bio-clinical parameters for each patient were assessed, including biomarkers associated with the CRAB criteria. The DNN model processed several indices, such as immunoglobulin, hemoglobin (HGB), serum creatinine, κ/λ light chain ratio, κ light chain, λ light chain, globulin, and immunofixation electrophoresis typing (IGA, G, M, κ, λ), along with CD markers (CD27, CD117, CD38, CD19, CD138). The mechanism-driven learning framework of the deep learning model is illustrated in Fig. 2.

Improved loss function

In multi-class classification tasks, deep learning models commonly utilize the cross-entropy loss function to evaluate predictive performance. This function ensures that the model’s outputs $f_y$ remain within the range of 0 to 1. During optimization, gradient descent is employed to minimize the cross-entropy loss, guiding the model towards its optimal parameter configuration ${\mathrm{f}}_{\mathrm{y}}\left(\mathrm{x}\right)$.

(4) $$\mathrm{l}\left(\mathrm{y},\mathrm{f}\left(\mathrm{x}\right)\right)=-\mathrm{l}\mathrm{o}\mathrm{g}{\displaystyle \frac{{\mathrm{e}}^{{\mathrm{f}}_{\mathrm{y}}\left(\mathrm{x}\right)}}{{\sum }_{{\mathrm{y}}^{\mathrm{\prime }}\in \left[\mathrm{L}\right]}{\mathrm{e}}^{{\mathrm{f}}_{{\mathrm{y}}^{\mathrm{\prime }}}\left(\mathrm{x}\right)}}=\mathrm{l}\mathrm{o}\mathrm{g}\left[1+{\sum }_{{\mathrm{y}}^{\mathrm{\prime }}\ne \mathrm{y}}{\mathrm{e}}^{\left({\mathrm{f}}_{{\mathrm{y}}^{\mathrm{\prime }}}\left(\mathrm{x}\right)-{\mathrm{f}}_{\mathrm{y}}\left(\mathrm{x}\right)\right)}\right].}$$

In predicting MM, the dataset contains a disproportionately higher number of MM cases. To enhance the model’s accuracy for rarer conditions, such as MGUS, a Logit adjustment was applied to the loss function (Menon et al., 2020). In this adjustment, ${\pi }_{\mathrm{y}}^{\mathrm{\prime }}$ represents the frequency of each class, $f_y$ denotes the model’s output, $\tau$ is a weighting factor, and ${\mathrm{f}}_{\mathrm{y}}(\mathrm{x})+\tau \log {\pi }_{\mathrm{y}}$ ensures that the model’s output distribution more closely aligns with the true data distribution. As shown in the formula, Logit adjustment proportionally modifies the spatial distribution of ${\mathrm{e}}^{\left({\mathrm{f}}_{{\mathrm{y}}^{\prime }}\left(\mathrm{x}\right)-{\mathrm{f}}_{\mathrm{y}}\left(\mathrm{x}\right)\right)}$ based on the ratio of $\frac{{\mathrm{\pi }}_{{\mathrm{y}}^{\prime }}}{{\mathrm{\pi }}_{\mathrm{y}}}$, thereby increasing the margin between ${\mathrm{f}}_{\mathrm{y}}^{\mathrm{\prime }}$ and ${\mathrm{f}}_{\mathrm{y}}$ and assigning greater weight to underrepresented data.

(5) $$l\left(y,f\left(x\right)\right)=-log{\displaystyle \frac{e^{f_y\left(x\right)+\tau log{\pi }_y}}{{\sum }_{y^{\mathrm{\prime }}\in \left[L\right]}{e}^{f_{y^{\mathrm{\prime }}}\left(x\right)+\tau log{\pi }_y\mathrm{\prime }}}=log\left[1+{\sum }_{y\mathrm{\prime }\ne y}\left({\displaystyle \frac{{\pi }_{y\mathrm{\prime }}}{{\pi }_y}}\right)\cdot e^{\left(f_{y^{\mathrm{\prime }}}\left(x\right)-f_y\left(x\right)\right)}\right].}$$

Building on this foundation, the deep learning model was configured with a learning rate of 0.0001 and a batch size of 128. This setup ultimately enabled the successful classification and differentiation among a-MM, light chain MM, MGUS, and LPL within a dataset characterized by well-defined features.

Performance evaluation

1. Evaluation Indexes of a Diagnostic Model

Precision, recall, and F1 scores are standard metrics used to evaluate the performance of machine learning algorithms. These metrics are defined by the following equations.

(6) $$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\left(\mathrm{P}\right)={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(7) $$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}\left(\mathrm{R}\right)={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(8) $$\mathrm{F}1\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}={\displaystyle \frac{2\times \mathrm{P}\times \mathrm{R}}{\mathrm{P}+\mathrm{R}}.}$$

Precision (P) measures the accuracy of correctly identifying plasma cell dyscrasias, while recall (R) reflects the proportion of true cases successfully detected. The F1 score balances precision and recall, with higher values indicating superior model performance. True positives (TP) represent correct classifications of plasma cell dyscrasias, false positives (FP) refer to instances mistakenly classified as dyscrasias, and false negatives (FN) are actual cases that were missed by the model. These metrics apply similarly to non-plasma cell dyscrasias.

2. Evaluation indexes of the subtype classification model

Performance evaluation utilized sensitivity (Se), specificity (Sp), positive predictive value (PPV), and negative predictive value (NPV), all calculated from the standard confusion matrix.

(9) $$\mathrm{S}\mathrm{e}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(10) $$\mathrm{S}\mathrm{p}={\displaystyle \frac{\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}}}$$

(11) $$\mathrm{P}\mathrm{P}\mathrm{V}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(12) $$\mathrm{N}\mathrm{P}\mathrm{V}={\displaystyle \frac{\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N}}.}$$

True positive, false positive, false negative, and true negative are represented as TP, FP, FN, and TN, respectively. In subtype classification models, these metrics are employed to evaluate various dimensions of model performance, offering insights into accuracy, precision, and effectiveness across different categories. By collectively considering these indices—sensitivity, specificity, PPV, and NPV—a comprehensive understanding of the model’s overall performance and suitability can be achieved. These metrics are essential for assessing the effectiveness of subtype classification models, providing valuable information on the model’s performance under different conditions.

All experiments referenced in this paper were conducted using Python version 3.10. The SVM and decision tree models were implemented through the scikit-learn library, while deep learning algorithms were developed using the PyTorch framework. The computational hardware used included an Intel i5-11400 CPU and an NVIDIA GeForce RTX 3060 GPU.