A method and system for predicting transformation of follicular lymphoma in a non-invasive whole-body assessment

Abstract

This invention discloses a non-invasive systemic assessment method and system for predicting the transformation of follicular lymphoma. The method includes the following steps: S1, obtaining four characteristics of the subject: peripheral blood hepatitis B surface antigen (HBsAg), neutrophil-to-lymphocyte ratio (NLR), serum lactate dehydrogenase (LDH), and PET-CT reporting maximum standardized uptake (SUVmax); S2, determining the weight of each characteristic based on its importance, with the weights from high to low being HBsAg, LDH, NLR, and SUVmax, and converting each characteristic into a risk score of 0-3 points using a unique fixed node; S3, weighted summing of the scores of the four characteristics to obtain a total score S, establishing a correspondence between the total score and transformation, constructing a web-based calculator and transformation scoring system, and outputting the individual transformation probability prediction result. By embedding the four conventional indicators (HBsAg, NLR, LDH, and SUVmax) selected by LASSO and Boruta into a lightweight CatBoost model, the transformation probability can be output, realizing a non-invasive assessment of the histological transformation risk of follicular lymphoma at initial diagnosis.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis technology, specifically to a non-invasive whole-body assessment method and system for predicting transformed follicular lymphoma. Background Technology

[0002] Follicular lymphoma (FL) is a common, indolent non-Hodgkin lymphoma, accounting for approximately 20-30% of all non-Hodgkin lymphomas. Although most patients have an indolent disease course, with a median survival of nearly 20 years and a generally good prognosis, some patients undergo histological transformation (HT) during the disease's progression, developing into a more aggressive lymphoma (primarily diffuse large B-cell lymphoma), known as transformed follicular lymphoma (t-FL). Statistics show that the annual incidence of HT in FL patients after diagnosis is approximately 1%. Once transformation occurs, the patient's treatment response declines sharply, with significantly shortened progression-free survival (PFS) and overall survival (OS). The median survival after transformation is only 5 years, becoming a key adverse event affecting the prognosis of FL patients. Therefore, early identification of patients at high risk of transformation and individualized adjustments to follow-up or treatment strategies have become an urgent need in hematology clinicians. Currently, domestic and international guidelines still rely on clinical scoring systems such as FLIPI and FLIPI-2 as the primary prognostic tools. However, these models only achieve a predictive accuracy of 0.60-0.65 and cannot indicate "when" or "which" patient is about to transform. In recent years, research on molecular biomarkers based on high-throughput sequencing or gene chips has gradually increased, attempting to supplement the shortcomings of traditional scoring systems by exploring differences in the expression of genes related to the tumor microenvironment, apoptosis pathways, or B-cell receptor signaling.

[0003] For example, there is a US patent with publication number US20090092991A1, which relates to a kit, method, and computer system for predicting 5-year survival outcomes in follicular lymphoma using gene pair expression levels. However, the US patent with publication number US20090092991A1 has the following problems: First, the detection subjects are limited to fresh or paraffin-embedded tumor tissue, requiring RNA extraction and microarray or qRT-PCR, which is complex, costly, and difficult to replicate in outpatient follow-ups; Second, the model input consists of dozens of mRNA expression values, which places extremely high demands on laboratory quality control, data normalization, and bioinformatics support, making it difficult to implement in primary hospitals; Third, the predicted endpoint is "5-year overall survival," without specifically targeting the clinically critical event of "histological transformation," and no stratification threshold is provided to guide the frequency of follow-ups; Fourth, the algorithm relies on chip platforms and multi-step log transformation, lacking a lightweight solution that can be computed at the bedside, resulting in long reporting cycles and poor patient compliance. Therefore, clinical practice still needs a predictive tool that is non-invasive in obtaining samples, uses conventional indicators, is quick to calculate, and can provide early warnings of individual conversion risks in order to overcome the limitations of existing gene expression protocols. Summary of the Invention

[0004] This invention proposes a non-invasive systemic assessment method and system for predicting the transformation risk of follicular lymphoma. By embedding four conventional indicators selected by LASSO and Boruta (HBsAg, NLR, LDH, and SUVmax) into a lightweight CatBoost model, and mapping the total score to the transformation probability in real time, the method achieves non-invasive assessment of the histological transformation risk of follicular lymphoma at initial diagnosis.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a non-invasive systemic assessment method for predicting transformed follicular lymphoma, comprising the following steps: S1. Obtain four indicators from the subjects: peripheral blood hepatitis B surface antigen HBsAg, neutrophil to lymphocyte ratio NLR, serum lactate dehydrogenase LDH, and PET-CT reporting maximum standard uptake value SUVmax. S2, based on the SHAP average value, determine the weights of each feature from high to low as HBsAg, SUVmax, LDH, NLR, and convert each feature into a risk score of 0-3 points according to a unique fixed node; S3, the total score S is obtained by weighted summing of the scores of the four features, and is mapped to the conversion probability P within 24 months using the Logistic formula.

[0006] This technical solution discovers the correlation between host hepatitis B infection status (HBsAg) and histological transformation of follicular lymphoma, and based on this, constructs a multidimensional non-invasive prediction system, breaking through the prognostic assessment framework based solely on tumor burden.

[0007] Preferably, step S1 includes: S11: Obtain clinical data of patients with follicular lymphoma, preset inclusion and exclusion criteria, and obtain input data; S12, the input data is divided into a conversion queue and a non-conversion queue based on whether the patients have been converted. Each queue is randomly divided into a training queue and a validation queue in a 7:3 ratio. S13. In the training queue, LASSO and Boruta were used to screen important features related to the research outcome. The intersection features HBsAg, NLR, LDH and SUVmax obtained by LASSO and Boruta were identified as important feature factors related to transformation.

[0008] Preferably, based on four metrics, Logistic Regression, SVM, GBM, Neural Network, Xgboost, KNN, Adaboost, and CatBoost models are developed and trained. The predictive performance of the models in the training and validation queues is evaluated by ROC curves, and the GBM model with high predictive performance is selected as the prediction model.

[0009] Preferably, in step S2, converting each feature into a risk score of 0-3 points according to a unique fixed node includes: 0 points for HBsAg negative and 3 points for positive; 0 points for SUVmax < 11, 1 point for SUVmax between 11 and 16, and 2 points for SUVmax > 16; 0 points for LDH < 198 U / L, 1 point for LDH between 198 and 275 U / L, and 2 points for LDH > 275 U / L; 0 points for NLR < 2.8, 1 point for NLR between 2.8 and 3.2, and 2 points for NLR > 3.2.

[0010] Preferably, step S3 includes: developing a scoring table based on the importance of the four features and the clinical cutoff point, and weighting and accumulating the scores of the four features according to the scoring table to obtain a total score.

[0011] Preferably, step S3 includes: classifying risk levels based on the total score S, with S of 0-3 being the low-risk group, S of 4-6 being the medium-risk group, and S of 7-9 being the high-risk group.

[0012] Preferably, in step S3, the clinical cutoff point is determined based on the median or mean of the clinical data.

[0013] Preferably, the pre-defined inclusion criteria include: initial diagnosis of primary FL; no history of other tumors; and availability of comprehensive medical records.

[0014] Preferably, the pre-defined inclusion criteria include: initial diagnosis of primary FL; no history of other tumors; and availability of comprehensive medical records.

[0015] Preferably, the pre-defined exclusion criteria include: a history of other tumors; incomplete medical records; histopathological diagnosis of FL; transformation from FL to other histological types; and patients initially diagnosed with complex lymphoma.

[0016] The present invention also adopts the following technical solution: a non-invasive whole-body assessment system for predicting transformed follicular lymphoma, comprising an input module for reading or manually inputting HBsAg, NLR, LDH, and SUVmax; a processor connected to the input module, with built-in fixed nodes and a Logistic formula; and a display module connected to the processor, outputting risk grouping and transformation probability. When the processor executes the system, it implements the non-invasive whole-body assessment method for predicting transformed follicular lymphoma as described in any one of claims 1-9.

[0017] The beneficial effects of this invention are: (1) This invention innovatively incorporates HBsAg and NLR into the transformation prediction and evaluation system, revealing the relationship between HBsAg, NLR and transformation, and providing more comprehensive prognostic information; (2) The web calculator and conversion scoring system constructed in this invention have good clinical applicability. Doctors can directly input patient data to obtain individualized conversion probability predictions, providing a scientific basis for clinical decision-making. (3) The reliability of the model was confirmed through a multi-dimensional validation system. Decision curve analysis showed that the method of the present invention has higher clinical net benefits than the FLIPI score and FLIPI-2 score. Attached Figure Description

[0018] Figure 1 This is a flowchart of the present invention.

[0019] Figure 2 This is the ROC curve of the training queue in Embodiment 1 of the present invention.

[0020] Figure 3 This is the ROC curve of the verification queue in Embodiment 1 of the present invention.

[0021] Figure 4 This is the SHAP diagram of the GBM model in Embodiment 1 of the present invention.

[0022] Figure 5 This is a schematic diagram of the prediction result output in Embodiment 1 of the present invention.

[0023] Figure 6 This is a comparison chart of the ROC curves of the scoring system of this invention with the FLIPI and FLPI2 scores.

[0024] Figure 7 This is a comparison chart of the DCA curves of the scoring system of this invention with those of FLIPI and FLPI2. Detailed Implementation

[0025] This embodiment provides a non-invasive, systemic assessment method for predicting transformed follicular lymphoma, such as... Figure 1 As shown.

[0026] First, clinical data from patients with follicular lymphoma were obtained. The research was based on newly diagnosed follicular lymphoma patients who visited a certain hospital consecutively from January 2015 to December 2023.

[0027] Clinical data collection was conducted strictly in accordance with pre-set standards.

[0028] Inclusion criteria include: (1) Primary FL diagnosed at the initial diagnosis according to the World Health Organization Classification of Tumors of Hematopoietic and Lymphoid Tissue (2022 edition), with histological grade I-II or IIIa; (2) No history of malignant tumors (excluding malignant tumors with good prognosis and no recurrence for 5 years after cure, such as basal cell carcinoma of the skin and cervical carcinoma in situ); (3) Complete medical records, including at least baseline demographic characteristics, pathological diagnosis report, immunohistochemical results, imaging assessment data and initial treatment plan. Exclusion criteria include: (1) History of other malignant tumors (excluding tumors that meet the above exceptions); (2) Incomplete medical records, missing key information (such as pathological grade and follow-up data) and cannot be obtained through supplementary medical record search; (3) Histological diagnosis of FL grade IIIb; (4) Transformed into other types of lymphoma at the initial diagnosis (such as Burkitt lymphoma, Hodgkin lymphoma, etc.); (5) Initial diagnosis of complex lymphoma (such as FL combined with diffuse large B-cell lymphoma).

[0029] Ultimately, 357 patients met the study criteria, of whom 19 underwent histological transformation during follow-up. The median follow-up time was 46.7 months, and the transformation rate during follow-up was 5.3%.

[0030] Blood biochemical parameters were collected during the patient's initial visit, and the tests were completed within 12 hours after fasting venous blood collection. The lymphocyte-monocyte ratio (LMR) was calculated using the formula: LMR = L / M, where L is the absolute lymphocyte count and M is the absolute monocyte count. The neutrophil-lymphocyte ratio (NLR) was calculated using the formula: NLR = N / L, where N is the absolute neutrophil count and L is the absolute lymphocyte count.

[0031] All cases were pathologically confirmed as grade I-IIIA and had no history of other malignant tumors. After collecting peripheral blood and PET-CT data, the overall cohort was randomly divided into training and validation parts in a 7:3 ratio. In the training subset, LASSO regression and Boruta dual screening were performed sequentially. The two algorithms consistently pointed to four indicators, HBsAg, NLR, LDH and SUVmax, at their intersection, thus establishing the core variable pool for subsequent modeling. Building upon this foundation, we further constructed eight machine learning models, including Logistic Regression, SVM, GBM, Neural Network, XGBoost, KNN, Adaboost, and CatBoost. Hyperparameters were determined using grid search combined with manual fine-tuning. Predictive performance was then compared horizontally within the validation subset using ROC curves. The results showed that the CatBoost model had the best predictive performance, with an AUC of 0.939 (95% CI: 0.895–0.984) on the training set and 0.863 (95% CI: 0.660–1.000) on the validation set, significantly outperforming the other candidate algorithms. Therefore, it was selected as the final prediction engine.

[0032] Specifically, in this embodiment, the derivation queue is randomly divided into two subsets using the random number method: 70% of the patients are in the training queue, and the remaining 30% are in the internal validation queue, which are used to build and screen the best prediction model.

[0033] LASSO regression and the Boruta algorithm were used in the training cohort to screen for important features associated with the study outcome.

[0034] LASSO regression can reduce the risk of overfitting by using a compression coefficient. It screens out 16 features, including HBsAg, NLR, LDH, SUVmax, Ann Arbor stage, HBV DNA, FLIPI-2 score, number of affected lymph nodes, extranodal involvement, immunohistochemical BCL2, maximum lymph node diameter, β2 microglobulin, spleen involvement, hemoglobin, lymphocyte-to-monocyte ratio (LMR), and platelets.

[0035] The Boruta algorithm compares artificially created features (shadow features) with the original features to determine the importance of the original features, thereby selecting features that are truly relevant to the target variable. It selected six features: hepatitis B surface antigen (HBsAg), neutrophil-to-lymphocyte ratio (NLR), lactate dehydrogenase (LDH), maximum standardized uptake value (SUVmax), lymphocyte count, and histological grade.

[0036] Finally, by taking into account the intersection of the features selected by the two methods and clinical relevance, we ultimately selected four features for the machine learning analysis. These features include HBsAg, SUVmax, LDH, and NLR.

[0037] The four features selected at the end were used to develop machine learning prediction models, resulting in eight machine learning models: Logistic Regression, SVM, GBM, Neural Network, Xgboost, KNN, Adaboost, and CatBoost.

[0038] To optimize the prediction model, a grid search combined with manual fine-tuning was used to obtain the final hyperparameters.

[0039] Eight machine learning models were trained on the training and validation sets. The predictive performance of the models on the training and validation sets was evaluated by ROC curves. The AUC value reflects the model's ability to distinguish between positive and negative examples. The closer the value is to 1, the better the performance. The 95% confidence interval was also marked to reflect the reliability of the results.

[0040] The ROC of the eight ML models on the training and test sets are as follows: Figure 2 and Figure 3 As shown in the diagram, in the training set, the GBM model (green) has the highest AUC (0.999, 95% CI: 0.996-1.000), exhibiting the best predictive performance. The KNN model (blue) has an AUC of 0.977 (95% CI: 0.960-0.994), and the CatBoost model (pink) has an AUC of 0.939 (95% CI: 0.895-0.984), showing slightly lower performance. Models such as SVM and Adaboost have AUCs between 0.765 and 0.783, indicating moderate performance. In the internal validation set, the CatBoost model (pink) had an AUC of 0.863 (95% CI: 0.660-1.000), and the GBM model (green) had an AUC of 0.853 (95% CI: 0.698-1.000), maintaining good performance. Models like KNN and Adaboost had AUCs between 0.782 and 0.786, indicating acceptable stability. SVM and Neural Network models had AUCs < 0.5, showing poor performance. Ultimately, the CatBoost model was determined to have the best predictive performance.

[0041] To transform the aforementioned black-box model into a clinically bedside-operable tool, this embodiment uses the SHAP mean absolute value to interpret CatBoost. The features are sorted by importance from highest to lowest as follows: HBsAg, LDH, NLR, and SUVmax. Weights are assigned appropriately for clinical application. The specific scoring rules are as follows: SUVmax is divided between 11 and 16; LDH between 198 U / L and 275 U / L; NLR between 2.8 and 3.2; and HBsAg retains its negative / positive binary attribute. Corresponding risk scores are 0 / 1 / 2 and 0 / 3, respectively, thus forming the HT score table.

[0042] Any new patient can obtain scores for each factor within ten seconds by simply comparing the four immediately available measured values ​​to a table. These scores are then summed to form a total score S, ranging from 0 to 9. In this study, scores of 0-3, 4-6, and 7-9 were defined as low, intermediate, and high risk levels, respectively, indicating different risk stratifications for transformation of follicular lymphoma. A clear stratification trend was observed in progression-free survival and overall survival, and corresponding follow-up intervals of 12 months, 6 months, and 3 months were recommended to facilitate immediate decision-making by clinicians.

[0043] Specifically, in this embodiment, the weight allocation of each feature is determined based on the importance of variables in the CatBoost prediction model. To intuitively illustrate the role of each feature in the CatBoost model, the importance of features and their impact on prediction are analyzed through the mean absolute SHAP value and SHAP scatter plot.

[0044] First, based on the feature importance bar chart, HBsAg has the greatest impact on the model prediction, followed by LDH, NLR, and SUVmax, which have relatively the smallest impact (the longer the average SHAP value, the more significant the feature's impact on the model).

[0045] Furthermore, CatBoost's SHAP plot further illustrates the correlation between the "level" of each feature's value and its prediction contribution, such as... Figure 4 As shown, the vertical axis represents the four features HBsAg, SUVmax, LDH, and NLR, while the horizontal axis represents the SHAP value. A positive SHAP value increases the probability of predicting "positive examples," while a negative value decreases it. In the color bar on the right, orange represents high feature values ​​and purple represents low feature values.

[0046] When HBsAg values ​​are high (orange dots), they are mostly distributed in the positive SHAP region, indicating that high values ​​are more likely to increase the probability of the model predicting "positive examples". High SUVmax and LDH values ​​(orange dots) also tend to cluster in the positive SHAP region, suggesting that increasing these indicators will have a positive impact on the prediction results. Although the relationship between NLR eigenvalues ​​and SHAP values ​​is more complex, the distribution of points still shows the difference in the contribution of different values ​​to the model prediction.

[0047] In the CatBoost prediction model, the weights of the four features, from highest to lowest, are: HBsAg, LDH, NLR, and SUVmax. A new conversion scoring system is constructed based on these four features. The continuous variable is divided according to its mean value in converted and non-converted patients. Each feature is scored based on the magnitude of the SHAP mean. In this embodiment, HBsAg negative receives 0 points, positive receives 3 points; SUVmax < 11 receives 0 points, 11 - 16 receives 1 point, and > 16 receives 2 points; LDH: < 198 U / L receives 0 points, 198-275 U / L receives 1 point, and > 275 U / L receives 2 points; NLR: < 2.8 receives 0 points, 2.8-3.2 receives 1 point, and > 3.2 receives 2 points. After summing up the scores of each item, the risk level can be divided according to the total score into three risk groups: low, medium and high. Among them, 0-2 points are low risk group, 3-5 points are medium risk group, and 6-9 points are high risk group.

[0048] Specifically, in this embodiment, the simplified clinical score is converted into an individualized histological transformation risk probability through a web-based calculator, ensuring that the model maintains high predictive performance while possessing excellent clinical interpretability and ease of use. The operating logic of the web-based calculator will be explained in detail below.

[0049] At its core is a "dual-track prediction engine," which contains two closely related but independent prediction paths that work together to serve the final risk assessment. These paths are the precise prediction core (underlying engine) and the clinically friendly interface (conversion layer and output).

[0050] The core of precise prediction can be based on a fully trained CatBoost machine learning model. This model takes raw data of four core features (HBsAg status, LDH value, SUVmax value, and NLR value) as input and, through its complex internal decision tree ensemble algorithm, directly outputs a high-precision, continuous histological transformation prediction probability value (P_catboost, ranging from 0 to 1). This probability represents the model's most accurate estimate of the patient's risk.

[0051] The clinically friendly interface is a conversion mechanism built to overcome the difficulty for clinicians to directly understand the output of complex models. It maps the precise probabilities calculated by the CatBoost model to a risk framework defined by the simplified HT scoring system (0-9 points), and finally presents them to users in the form of more intuitive risk levels (low / medium / high) and probability ranges or typical values ​​corresponding to the scores.

[0052] The mapping relationship from total score to conversion probability is a fixed lookup relationship based on retrospective cohort data, established through statistical modeling and validation. The process of establishing this mapping relationship is explained in detail below.

[0053] First, data preparation and scoring were performed. Using all patient data from this study, the total HT score (S, integer, 0-9) was calculated for each patient according to the established HT scoring rules (HBsAg: 0 / 3 points; LDH, SUVmax, NLR: 0 / 1 / 2 points each). At the same time, the pre-trained final CatBoost model was used to calculate the four original features of the same batch of patients to obtain the direct prediction probability (P_catboost) of each patient.

[0054] Next, a statistical correlation between the total score and the predicted probability was established. All patients were grouped according to their HT total score (S), forming multiple subgroups from S=0 to S=9. For each total group, the median (or mean) of the predicted probability (P_catboost) obtained by the CatBoost model was calculated. This value represents the "typical" or "central" risk level of the population with that total score. At the same time, the distribution range of the predicted probability of the group (e.g., interquartile range: P25, P75) was calculated to reflect the risk variation within patients with the same score. This resulted in a one-to-one mapping table: each HT total score (S) corresponds to a typical predicted probability value (P_typical) and its distribution range confirmed by the precise model.

[0055] Then, the mapping relationship was validated and solidified. An internal consistency check was conducted to observe whether the mapping relationship met clinical expectations, that is, the higher the total score, the higher the typical predictive probability P_typical should be. The results of this study fully met this rule. Then, the Kaplan-Meier method was used to perform survival analysis (PFS, OS) by grouping with HT total score and by grouping with risk thresholds defined by P_typical. The validation results confirmed that the risk stratification based on HT total score and the risk stratification based on model probability have a high degree of consistency in prognostic ability, thus verifying the effectiveness of the mapping relationship in reverse. Finally, the validated "total score-typical probability / probability range" mapping table was integrated into the backend logic of the web calculator as a solidified data asset.

[0056] In the application phase of the real-time calculation process of the web calculator, after the clinician inputs the patient's four indicators into the web interface, the calculator first receives the HBsAg status (yes / no), LDH value, SUVmax value, and NLR value input by the user and performs preprocessing. Then, it performs parallel dual-track calculation: Track A (precise calculation) inputs the four raw data into the pre-loaded CatBoost model in the background to instantly calculate the accurate prediction probability P_catboost. Track B (simple scoring) converts the four inputs into HT scores according to preset rules and calculates the total score S.

[0057] In the results integration and output stage, the system first displays the intuitive HT total score S and the corresponding risk level (e.g., 0-3 points = low risk, 4-6 points = intermediate risk, 7-9 points = high risk). Then, it searches for the typical predicted probability range associated with the total score S based on the "total score-probability" mapping table fixed in the background (e.g., "Your score is 5 points (intermediate risk), and the typical histological transformation risk probability of patients with similar scores in the historical model is approximately XX% - YY%").

[0058] To further enhance personalization, the system can compare the patient's precise probability P_catboost calculated in track A with the typical probability range corresponding to their total score S. If P_catboost falls within the typical range, the consistency of the results is emphasized; if there is a significant deviation, prompts are given to guide doctors to focus on the specificity of the patient's characteristic combination, reflecting the depth of individualized assessment. The final webpage clearly displays the HT total score and risk level, the histological transformation risk probability estimate corresponding to that total score (based on historical model mapping), and optional personalized probability prompts.

[0059] In an internal validation cohort of 107 patients, the method achieved an AUC of 0.863, an improvement of 0.133 compared to the calculated FLIPI score (DeLong test p=0.0021). The decision curve showed an average net benefit increase of 0.028 within the 5-25% threshold range, equivalent to reducing false positive biopsies by 2.8 per 100 patients screened and simultaneously detecting 2.8 more true conversions. Since all input indicators are routine outpatient items and the calculation process only involves four arithmetic operations, this embodiment successfully compresses the machine learning model that originally relied on high-dimensional features and cloud computing power into a lightweight tool that can be printed on paper cards or embedded in an offline app, enabling non-invasive, real-time, and whole-body assessment in primary care settings, completing the transformation from "black box prediction" to "bedside usability".

[0060] To enhance the clinical applicability of this model, a web-based interactive dynamic bar chart application was developed using Shiny. This application predicts a patient's risk of histological transformation after inputting the actual values ​​of the four features required by the model. To visually illustrate the application of the risk prediction model for follicular lymphoma transformation, an example is given: a patient with "HBsAg positive, maximum standardized uptake (SUVmax) of 12.1, lactate dehydrogenase (LDH) of 255, and a neutrophil-to-lymphocyte ratio (NLR) of 2.63," demonstrating the predicted probability of follicular lymphoma transformation. The final output shows a 74.71% risk of follicular lymphoma transformation for this patient, indicating a high risk of transformation. Clinicians can use this prediction result to further develop follow-up or intervention plans, such as... Figure 5 As shown.

[0061] The clinical net benefit of the model was assessed using ROC and DCA curves and compared with the classic FLIPI score and FLIPI-2 score. The comparison is shown in the figure below. Figure 6 and Figure 7 As shown, the new scoring system outperforms the traditional FLIPI and FLPI2 scoring systems in predicting conversions.

[0062] A non-invasive, whole-body assessment system for predicting transformed follicular lymphoma enables primary healthcare institutions to achieve predictions with accuracy equivalent to the CatBoost model, even without cloud computing or high-configuration computers.

[0063] First, the overall system architecture and hardware deployment will be explained.

[0064] The system is implemented using a three-tier architecture: clinical terminal - edge computing node - cloud backup. The clinical terminal is a mobile cart integrating a 21.5-inch 4K touchscreen, barcode scanner, RFID reader, and thermal printer. Inside, it houses an industrial-grade ARM SoC (quad-core Cortex-A78@2.4 GHz, with an 8-TOPS NPU), and expands to a 32 GB RAM, 1 TB NVMe edge computing node via a PCIe interface, responsible for all model inference and result visualization. The cloud is used only for low-bandwidth backups and model version updates at night, and does not interfere with real-time diagnosis and treatment processes, ensuring independent operation even without an external network. The entire system is powered by a medical-grade isolated power supply, complies with EMC standards YY 0505-2012, and has an IP54 protection rating, making it plug-and-play suitable for various scenarios such as outpatient corridors, day chemotherapy centers, and mobile clinics in primary hospitals. After the system is powered on, it automatically performs a self-test to check the NPU computing power, memory ECC, storage lifespan, and printer status. Only when all indicators are green can the system enter the "pending test" interface to avoid failure at critical clinical moments.

[0065] The data acquisition module provides three entry points: ① in-hospital LIS / HIS interface, ② local Excel batch import, and ③ manual quick entry. For the interface mode, the system uses the HL7-FHIR standard to monitor laboratory and imaging events. Once it detects "pathological diagnosis code ICD-O 9690 / 3 (follicular lymphoma)" and "examination status = approved," it automatically retrieves the four required indicators. HBsAg is taken from the immunoluminescence report, NLR is calculated in real-time from the absolute values ​​of neutrophils and lymphocytes in the complete blood count, LDH comes from the biochemistry report, and SUVmax is extracted from the "SUVbw max" field in the PACS dicom header. If any indicator is missing, a red exclamation mark pops up on the interface, and the subsequent process is paused to prevent "inference with existing disease." For scenarios without an interface at the grassroots level, doctors can scan the patient's wristband barcode. The system automatically opens a "quick entry" pop-up window, presenting the four indicators in a stepped control format. The units are internationally recognized symbols, and the range validation rules are consistent with the thresholds in the instruction manual (e.g., NLR is not allowed to be >15). Illegal values ​​are immediately indicated by audio and visual prompts. After all fields are completed, the system generates a SHA-256 checksum, which is bound to the patient ID and timestamp and written into the local blockchain ledger to ensure that it is traceable and tamper-proof afterward.

[0066] Inside the edge node, all statistical parameters obtained during the training phase have been burned into the "fixed node" and do not need to be trained again. Specifically, it includes: 1) Standardized dictionary: HBsAg negative=0, positive=1; cut-off values ​​of SUVmax, LDH, and NLR are stored in EEPROM in uint16 format; 2) Scoring lookup table: four-dimensional array int8 score[2][3][3][3], with a total of 54 combinations, and direct table lookup takes 0.8 ms; 3) Logistic coefficient: float32 bias=-2.1, weight=0.38, using IEEE754 single precision, saving 50% of storage; 4) SHAP interpretation matrix: int8 shap[4]

[54] , used for subsequent result interpretation radar chart drawing.

[0067] During inference, the CPU first converts the four raw values ​​into discrete indices, then uses pointer offset to locate the total score S in one step, and finally calls the single instruction expf() of the ARM CMSIS-DSP library to complete the probability calculation. The entire pipeline ends within 1.2 ms, with a peak memory usage of no more than 256 KB. Even when facing a peak outpatient period of 300 patients per day, the CPU usage is still less than 15%.

[0068] The display module is redrawn using Qt for MCUs, ensuring readability even under direct sunlight at 600 nits. The results page is divided into three columns: The left side is the "probability dashboard," which uses SVG vector animation. The pointer smoothly rotates from 0° to the corresponding angle (e.g., 65.38% corresponds to 117°). Below, it displays "Probability of histological transformation within 24 months: 65.38%" in extra-large font, and gives the 95% CI (55.1%-74.2%) below the percentage to help doctors measure uncertainty. The middle column is the "Risk Grouping Card". Low, medium and high risk are represented by green, yellow and red respectively. The card provides follow-up recommendations, next examination items and estimated costs. For example, the high-risk group suggests "PET-CT follow-up + unilateral bone marrow biopsy within 6 weeks, with an estimated total cost of ¥3200". The right side is an "interpretive radar chart," with four axes corresponding to HBsAg, SUVmax, LDH, and NLR, respectively. The red-filled areas show the degree to which each variable contributes to the positive probability of the current patient. Hovering the mouse over the data will display the specific SHAP value, which meets the "interpretable" requirements of medical insurance and drug regulatory authorities for artificial intelligence.

[0069] If a doctor disagrees with the results, they can click the "Adjust" button. The system will then display four sliders, allowing manual fine-tuning of any indicator within a reasonable range. The probability and radar chart are linked in real time, facilitating sensitivity analysis or a second consultation.

[0070] After clicking "Generate Report", the system first calls AES-256 to encrypt the patient's sensitive information in the background, and then generates a PDF according to the "Electronic Medical Record Sharing Document Specification". The file name includes "Patient ID + Time + Version Number" and the content includes: the original four indicators, risk score S, conversion probability P, follow-up suggestions, and doctor's electronic signature area.

[0071] The report is simultaneously pushed to the EMR via HL7 ORU^R01 message and generates task lists in three queues: nursing station, day ward, and follow-up center. If a patient is classified as high-risk, the system will automatically send a WeChat message to the "Lymphoma MDT" WeChat group robot, containing the first letter of the patient's name in pinyin, the probability value, and an outpatient appointment link, achieving seamless multidisciplinary collaboration. Printing uses 80 mm thermal paper, with a 30-second output time. The printout can be pasted into the outpatient medical record and can also be viewed in color on a mobile device after scanning a QR code.

[0072] Considering the differences in the level of informatization among different hospitals, this system provides two configurations: "standalone offline version" and "interface version".

[0073] The standalone version does not require HIS integration; doctors can generate reports by scanning barcodes or manually entering four indicators, making it suitable for grassroots institutions that do not yet have an interface budget. The interface version supports the HL7-FHIR standard and can be integrated with HIS vendors to achieve automatic data reading.

[0074] In terms of operation and maintenance, a "Follicular Lymphoma Prediction System Technical Team" was established, responsible for: a) quarterly remote inspection of the health of the equipment's CPU, storage, and NPU; b) annual on-site calibration, including monitor colorimetry, printer paper feed accuracy, and system time synchronization; c) when updating model weights, providing a signed differential package, which the hospital's information department can complete the upgrade on its own during the night window, and can roll back with one click if the upgrade fails.

[0075] This invention innovatively incorporates HBsAg, SUVmax, LDH, and NLR into the conversion prediction and assessment system, revealing the relationship between these four key indicators and conversion, and providing more comprehensive prognostic information. The web-based calculator and conversion scoring system constructed in this invention have good clinical applicability; physicians can directly input patient data to obtain individualized conversion probability predictions, providing a scientific basis for clinical decision-making. The reliability of the model has been confirmed through a multi-dimensional validation system, and decision curve analysis shows that the method of this invention has higher clinical net benefits than traditional FLIPI and FLIPI-2 scores. The beneficial effects of this invention are described in detail below.

[0076] First, the prediction accuracy is significantly improved. In the internal validation cohort (n=107), the AUC of this system reached 0.863. Compared with the classic scoring model for follicular lymphoma, the prediction results for transformation show a larger area under the AUC curve and DCA curve, indicating a significant clinical benefit.

[0077] Second, it offers non-invasive, real-time, and comprehensive assessment. The system only requires four indicators routinely collected during the initial consultation: HBsAg, NLR, LDH, and SUVmax. No additional punctures or expensive molecular tests are needed. The entire process, from data scanning / entry to report display, takes less than 5 seconds, truly enabling "one-click" whole-body conversion risk assessment at the outpatient bedside. This addresses the pain points of traditional biopsy sampling bias and the inability to repeat the procedure.

[0078] Third, the model is interpretable and the results are actionable. The importance of variables is explained by the SHAP mean absolute value, and the black box CatBoost is further transformed into a 0-9 point visual score. Doctors can intuitively see the positive or negative contribution of each patient's four indicators to the probability. At the same time, the system provides specific follow-up intervals (12 months, 6 months, and 3 months) for low, medium, and high risks, realizing a closed loop of "prediction-decision-action" and avoiding over- or under-monitoring.

[0079] Fourth, it is lightweight and available offline. All algorithms and parameters are fixed on edge nodes, with a single chip power consumption of <8W, and no GPU data center is required. Even if primary hospitals have no external network or cloud resources, they can complete the same quality of assessment as tertiary hospitals in a single-machine environment, providing a standardized tool for hierarchical medical treatment.

Claims

1. A non-invasive, systemic assessment method for predicting transformed follicular lymphoma, characterized in that, Includes the following steps: S1. Obtain four characteristics of the subject: peripheral blood hepatitis B surface antigen HBsAg, neutrophil to lymphocyte ratio NLR, serum lactate dehydrogenase LDH, and PET-CT reporting maximum standard uptake value SUVmax. S2, determine the weight of each feature based on its importance, with the weights from high to low being HBsAg, LDH, NLR, and SUVmax, and convert each feature into a risk score of 0-3 points according to a unique fixed node; S3, the scores of the four features are weighted and accumulated to obtain the total score S, the correspondence between the total score and the conversion is established, the web page calculator and conversion scoring system are constructed, and the individual conversion probability prediction results are output.

2. The method for predicting transformed follicular lymphoma through non-invasive whole-body assessment according to claim 1, characterized in that, Step S1 includes: S11: Obtain clinical data of patients with follicular lymphoma, preset inclusion and exclusion criteria, and obtain input data; S12, randomly divide the input data into a training queue and a validation queue in a 7:3 ratio; S13. In the training queue, LASSO and Boruta were used to screen important features related to the research outcome. The intersection features HBsAg, NLR, LDH and SUVmax obtained by LASSO and Boruta were identified as important feature factors related to transformation.

3. The method for predicting transformed follicular lymphoma through non-invasive whole-body assessment according to claim 2, characterized in that, Based on four features, Logistic Regression, SVM, GBM, Neural Network, Xgboost, KNN, Adaboost, and CatBoost models were developed and trained. The predictive performance of the models in the training and validation queues was evaluated by ROC curves, and the CatBoost model with high predictive performance was selected as the prediction model.

4. The method for predicting transformed follicular lymphoma through non-invasive whole-body assessment according to claim 1, characterized in that, In step S2, each feature is converted into a risk score of 0-3 points according to a unique fixed node, including: 0 points for HBsAg negative and 3 points for positive; 0 points for SUVmax < 11, 1 point for SUVmax between 11 and 16, and 2 points for SUVmax > 16; 0 points for LDH < 198 U / L, 1 point for LDH between 198 and 275 U / L, and 2 points for LDH > 275 U / L; 0 points for NLR < 2.8, 1 point for NLR between 2.8 and 3.2, and 2 points for NLR > 3.

2.

5. A non-invasive systemic assessment method for predicting transformed follicular lymphoma according to claim 1 or 2, characterized in that, Step S3 includes: developing a scoring table based on the importance of the four features and the clinical cutoff point, and weighting and accumulating the scores of the four features according to the scoring table to obtain the total score.

6. A non-invasive systemic assessment method for predicting transformed follicular lymphoma according to claim 1 or 2, characterized in that, Step S3 includes: classifying risk levels based on the total score S, with S of 0-3 being the low-risk group, S of 4-6 being the medium-risk group, and S of 7-9 being the high-risk group.

7. The method for predicting transformed follicular lymphoma through non-invasive whole-body assessment according to claim 5, characterized in that, In step S3, the clinical cutoff point is determined based on the median or mean of the clinical data.

8. The method for predicting transformed follicular lymphoma through non-invasive whole-body assessment according to claim 2, characterized in that, The pre-defined inclusion criteria included: initial diagnosis of primary FL; no history of other tumors; and availability of comprehensive medical records.

9. A non-invasive systemic assessment method for predicting transformed follicular lymphoma according to claim 2, characterized in that, Pre-defined exclusion criteria include: a history of other tumors; incomplete medical records; histopathological diagnosis of FL; transformation from FL to other histological types; and patients initially diagnosed with complex lymphoma.

10. A non-invasive, systemic assessment system for predicting transformed follicular lymphoma, characterized in that, The system includes an input module for reading or manually inputting HBsAg, NLR, LDH, and SUVmax; a processor connected to the input module with a built-in fixed node; and a display module connected to the processor for outputting risk grouping and conversion probability. When the processor executes the system, it implements the non-invasive whole-body assessment method for predicting converted follicular lymphoma as described in any one of claims 1-9.

Images