Machine learning classification of inflammatory bowel disease activity using white blood cell subsets

Abstract

Objective

The lack of a rapid, validated, consistent test for tracking disease activity in patients with inflammatory bowel disease (IBD) is currently a major challenge. Currently used biomarkers have notable disadvantages, such as the slow processing (faecal calprotectin) and the lack of specificity (bloodwork). White blood cell (WBC) subsets, also known as ‘the differential’, are commonly obtained in evaluating IBD patients, but there is minimal evidence on how these subsets relate to disease activity. Given the interplay between immune cells, it is possible that complex patterns in WBC subsets could be used to classify IBD activity. Machine learning (ML) could be used to reveal these changes. The aim of this study was to classify IBD activity via routine bloodwork results, using an ML approach.

Methods

1458 bloodwork measurements from 108 IBD patients were included in this analysis. Disease activity was classified by physician’s global assessment score. Four ML models were trained to classify active disease or remission based on routine bloodwork metrics (complete blood count, differential, albumin, erythrocyte sedimentation rate and C reactive protein).

Results

The optimal model, extreme gradient boosted decision trees, achieved a receiver operator characteristic area under the curve of 0.882. Feature analysis identified neutrophils, C reactive protein and albumin as consistently important contributors to the models. Additionally, no single individual biomarker was comparable to the ML model, and medications had only a minor impact on the ML model.

Conclusion

Classification of IBD activity can be augmented using ML analysis of commonly measured bloodwork parameters to help inform treatment plans and to improve IBD patient outcomes.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• The lack of a rapid, validated, consistent test for tracking disease activity in patients with inflammatory bowel disease (IBD) is currently a major challenge.

WHAT THIS STUDY ADDS

• This study found that machine learning models were able to classify IBD activity based on routinely collected, non-invasive biomarkers. The extreme gradient boosted decision trees model was statistically most robust with a receiver operating characteristic-area under the curve of 0.882.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• In the future, machine learning models could be applied to clinical practice to reduce the need for repeat colonoscopies and invasive procedures in IBD patients.

Introduction

Inflammatory bowel disease (IBD) is a chronic, autoinflammatory condition that has profound effects on the gastrointestinal tract. Classifying IBD activity is often very challenging. Activity of IBD is monitored using a variety of biomarkers and tools, all of which have substantial limitations.¹ Serum and stool biomarkers such as C reactive protein and faecal calprotectin are currently the most used in practice to supplement patient-reported symptoms, but all these metrics are either non-specific or limited by processing time. Additionally, symptoms and signs poorly reflect disease activity, especially in children.¹

Machine learning (ML) models have the benefit of being able to rapidly assess large amounts of complex data with multiple variables.² ML has previously been used to predict disease risk, diagnose patients, monitor treatment response and classify disease subtypes in a range of autoimmune diseases, including psoriasis, rheumatoid arthritis, systemic lupus erythematosus and multiple sclerosis³; but has yet to be applied to IBD activity classification. Recently, ML models have assisted with the diagnosis of IBD through uncovering complex differences in histology, endoscopic images, the faecal microbiota, electronic medical health record data and serological markers.⁴ These ML models show promise for use in diagnosing IBD,⁴ since it is a complex and multifaceted disease. However, once a patient is diagnosed and treated, it would be helpful if ML could be used to objectively classify a patient as in remission or as having ongoing active disease. Gold standard evaluations of disease activity are invasive (endoscopies) or costly (cross-sectional imaging); therefore, using ML to leverage data from non-invasive biomarkers would be ideal.⁴ Thus far, ML-based tools for classifying IBD activity using non-invasive biomarker analysis have not yet been developed.

The aim of this study was to investigate the ability of ML models to objectively classify IBD patient disease activity using routinely collected objective parameters. Changes in the white blood cell (WBC) differential were a focus of our models since these parameters are very commonly measured but lack any specific guidance as to how they should be interpreted in this context. This study also examines inter-individual differences in biomarkers between times of active disease and remission, which could drive future personalisation of IBD care, through the use of patient-specific ML models to classify disease activity.

Methods

Study population

Data were collected from patients with IBD in a regional IBD database at the Kingston Health Sciences Centre in Kingston, Ontario, Canada. A Strengthening the Reporting of Observational Studies in Epidemiology checklist is provided in online supplemental material. Patient involvement was not a part of this study. Clinical and laboratory data were extracted from the medical record for both adult and paediatric IBD cases identified in clinics between February 2007 and September 2025. All available data up to 3 months prior to the initial presentation of IBD were also included. To be included in the study, patients needed a diagnosis of IBD, and at least one bloodwork report with a WBC differential and a concurrent record of their disease activity within the same encounter. IBD activity was defined using the ImproveCareNow guidelines for the Physician Global Assessment Score (PGAS).⁵ This score provides a standardised assessment scale to rate the disease activity as in remission, mild, moderate or severely active, according to eight parameters. Encounters where there was another significant medical comorbidity at the time, such as malignancy or infection, were excluded. Encounters with no bloodwork results, or with an unknown disease activity state, were removed.

Data collection and preprocessing

There were 1458 encounters across 108 patients that met the inclusion criteria. The variables collected were sex, age, PGAS, IBD subtype, medication, faecal calprotectin, WBC differential (neutrophils, lymphocytes, monocytes, eosinophils, basophils), haemoglobin, haematocrit, platelets, albumin, erythrocyte sedimentation rate (ESR) and C reactive protein (CRP). The WBC differential and platelets were expressed in absolute count ($\times {10}^9$ cells/L). Categorical variables were sex (female, male or other), PGAS (remission, mild, moderate or severe), IBD subtype (Crohn’s disease (CD), ulcerative colitis (UC) or IBD-unclassified (IBD-U)) and medication (acetylsalicylic acid, infliximab, adalimumab, prednisone, azathioprine, vedolizumab, methotrexate, budesonide or ustekinumab). Faecal calprotectin was reported in micrograms per gram (µg/g). Haemoglobin and albumin were reported in grams per litre (g/L). Haematocrit was reported as the decimal proportion of total blood volume made up of red blood cells. ESR rate was reported in units of millimetres per hour (mm/hour). CRP was reported in units of milligrams per litre (mg/L).

All analysis and data cleaning were performed in R (V.4.4.1) using the tidyverse package (V.2.0.0). Graphing was performed using the ggplot2 package (V.3.5.1). Medications where there were less than four patients were removed from the medication analysis. Summary statistics were generated for the dataset using the gtsummary (V.2.0.2) package.

To satisfy principal component analysis (PCA) and some ML model requirements, standardisation was performed on numeric variables. Variables with a high percentage of missing values were omitted to avoid introducing bias when imputing missing values. The variables omitted were faecal calprotectin (87% missing), albumin (43% missing) and ESR (62% missing). Remaining missing values were imputed using the multivariate imputation by chained equations method with predictive mean matching, using the mice package (V.3.17.0).

Remaining numeric variables were tested for skewness using the e1071 (V.1.7.16) package, then transformed. Variables that were skewed left (skew<−0.5) were square root transformed. This included all five WBC differential variables, platelets and CRP. Variables that were right skewed (skew>0.5) were log₁₀ transformed with an offset of 0.1. This included haematocrit and haemoglobin. Standardisation was then performed on all numeric variables.

Dimensionality reduction and multicollinearity

PCA was performed on the standardised dataset to visually evaluate data structure. Haemoglobin and haematocrit are generally considered to represent similar clinical information and were strongly correlated in our dataset (online supplemental figure 1, R=0.96), using the ggpubr package (V.0.6.0) with Pearson’s correlation. Therefore, haemoglobin was removed from further analysis due to its higher percentage of missing values compared with haematocrit. No further variables were identified for removal when analysis was done for disease subtype and PGAS.

Machine learning

Four ML models were trained using the tidymodels package (V.1.3.0) following the workflow (figure 1). Models trained were random forest (RF), support vector machine (SVM), multilayer perceptron (MLP) and extreme gradient boosted decision trees (XGBoost). These models were chosen due to their range of model architectures. RF and XGBoost are both tree-based models, which are known for their ability to handle noisy data with missing values and non-normally distributed data, which is seen frequently in biomedical datasets. Additionally, tree-based models can provide feature importance scores, which aid in increasing the explainability of the models. SVM models handle non-linear relationships well, and artificial neural networks, like MLP, can readily uncover patterns in highly complex non-linear data. Specifically, the models used were rand_forest() with engine ‘ranger’, svm_rbf() with engine ‘kernlab’, mlp() with engine ‘nnet’ and boost_tree() with engine ‘xgboost’, respectively. All models trained were classification models.

Figure 1. Machine learning workflow. Steps in blue boxes were completed for all models. Steps in orange boxes were only followed for models specified in brackets. Numbers in brackets indicate the number of variables included in the standardised dataset compared with the variables in the non-standardised dataset. CV, cross-validation; ML, machine learning; MLP, multilayer perceptron; n, number of bloodwork sets; p, number of predictors; RF, random forest; SVM, support vector machine; XGBoost, extreme gradient boosted decision tree.

For all ML analysis, the data were split into an 80:20 training-test split, stratified by the activity variable. Data were standardised, as discussed above, for SVM and MLP models to fit model assumptions. Initial models were trained on the training set only and 10-fold cross-validated. The performance metrics calculated and reported for each model were accuracy, F₁-score, area under the receiver operating characteristic curve (ROC-AUC), precision, recall, sensitivity and specificity. Feature selection was performed for the RF model, where 20 features were retained to reduce computational cost while maintaining model performance. Model hyperparameters were then tuned with 10-fold cross-validation, including it as the resampling method.

Hyperparameters tuned for the RF model included the number of predictors sampled at each split (mtry), the number of trees in the forest (trees) and the minimum number of observations per terminal node (min_n). Hyperparameters tuned for the SVM model were cost, the regularisation parameter, which determines misclassification stringency, and rbf_sigma, which controls the spread of the Gaussian kernel. MLP hyperparameters tuned were hidden_units, penalty and epochs. hidden_units refers to the number of neurons in each hidden layer, penalty refers to ridge regression, a process that penalises large weights, and epochs refers to the number of training iterations. For the XGBoost model, the hyperparameters tuned were tree_depth, learn_rate, loss_reduction, min_n and mtry. tree_depth refers to the maximum depth of each tree, learn_rate controls how much each tree contributes to the model, loss_reduction controls the splitting of the nodes, min_n is the number of samples per leaf and mtry controls the number of predictors used for each tree.

Optimal hyperparameter combinations following tuning were selected based on the optimal F₁-score. Shapley additive explanations (SHAP) analysis was performed on the XGBoost model to visualise feature importances using the SHAPforxgboost package (V.0.1.3). Hyperparameter tuning was done only on the training set. Model performance was reported on analysis of the test set only, without any previous use of the test set data. Scripts used for analysis are available online (https://github.com/DanJMulder/diffsibdstudy).

Patient-specific statistical analysis

To investigate patient-specific variations in bloodwork between active disease and remission, patients with less than seven repeat bloodwork sets in both active disease and remission were removed from the initial dataset. 18 patients remained, with 501 bloodwork sets for this sub-set analysis.

For each patient included in this subset, a Welch’s t-test was performed to test between the means of each bloodwork value in active disease and remission. Resulting p values were adjusted using Bonferroni multiple testing correction. By ranking p values, the variable with the most significant difference in means between active disease and remission was identified for each patient. Of these variables, those that were significant at the p_Bonf<0.05 level were plotted over time to identify the variables that changed most between active disease and remission for each patient.

Impact of medication analysis

To investigate the effects of individual medications on patient bloodwork parameters, multi-way analyses of variance (MANOVAs) were conducted on the standardised dataset using the stats (V.3.6) package manova() to assess the effects of the different medications on WBC parameters. Resulting p values were adjusted with Bonferroni multiple testing correction. For medications that had a significant overall effect on bloodwork parameters, the results were further evaluated by individual patient to determine the effect on each WBC subtype on its own.

Results

Study population

Clinical characteristics of the cohort are summarised in table 1 and bloodwork and medications are summarised in online supplemental table 1. There were 64% of encounters from patients with active disease, and 36% were from patients in remission. The disease subtypes across the population were 69% with CD, 31% with UC and 0.5% with IBD-U. Male (42%) and female (58%) patients in the study population were relatively balanced. The median age of the patients was 34, with 432 bloodwork sets from paediatric patients (below 18 years old), and the remaining 1026 from adult patients.

Table 1. Cohort clinical characteristics.

Characteristic	Overall N=1458*	Active N=933*	Remission N=525*
Age	34 (1, 82)	34 (1, 76)	32 (3, 82)
Female	844 (58%)	542 (58%)	302 (58%)
Male	614 (42%)	391 (42%)	223 (42%)
Subtype
CD	999 (69%)	663 (71%)	336 (64%)
IBD-U	7 (0.5%)	7 (0.8%)	0 (0%)
UC	452 (31%)	263 (28%)	189 (36%)
PGAS
Mild	407 (28%)	407 (44%)	0 (0%)
Moderate	317 (22%)	317 (34%)	0 (0%)
Remission	525 (36%)	0 (0%)	525 (100%)
Severe	209 (14%)	209 (22%)	0 (0%)

^*Median (min, max); n (%).

CD, Crohn’s disease; IBD-U, IBD-unclassified; N, number of observations; PGAS, Physician Global Assessment Score; UC, ulcerative colitis.

Principal component analysis

The PCA scatterplot in figure 2A demonstrates the separation between disease activity states in the bloodwork data. In both scatterplots, the proportion of variance illustrated by the first two PCs is slightly above 39%. Notably, the degree of severity of activity is captured in the bloodwork, given that the separation from remission to mild to moderate to severe activity is stepwise, spanning the first two PCs (figure 2B).

Figure 2. (A) Principal component analysis scatterplot of disease activity. 95% CI for each group is shown by the coloured ellipses. The number of bloodwork sets in each group is noted in brackets in the legend. (B) PCA scatterplot of the Physician Global Assessment Score (PGAS). 95% CI for each group is shown by the coloured ellipses. The number of bloodwork sets in each group is noted in brackets in the legend. (C) Highest contributing variables to the first four principal components (PCs) from principal component analysis (PCA). Variables are ranked based on their relative contribution, with the top eight variables from each PC included. Sign of contribution of the variable is indicated by blue (negative) or red (positive).

Principal components (PCs) were also used to identify variables that contribute together to disease activity classification. The proportion of variance is spread among the PCs, with no single variable able to highly explain a significant portion of the variance in the data, thus justifying an ML approach. In PC1, the strongest contributing variables were platelets, monocytes, lymphocytes, eosinophils and basophils (figure 2C). Increased values in these variables increased the probability that the patient had active disease at that encounter. In PC2, increases in platelets, neutrophils and CRP all increased the likelihood the patient had active disease, while higher lymphocytes and haematocrit decreased the likelihood of active disease. All contributing variables to the first 5 PCs shown from the PCA plots in online supplemental figure 2A. Disease subtype distribution across our dataset was highly overlapping (online supplemental figure 2B), justifying analysis of these subtypes together.

Machine learning

Performance metrics for the final tuned model when evaluated on the test set data are shown in table 2. The XGBoost model had the highest performance for three metrics: accuracy (81.5%), F₁-score (0.859) and precision (0.841). The RF model had the highest recall (0.888) and ROC-AUC (0.898, online supplemental figure 3) of all models. MLP had approximately the same precision as the XGBoost model (0.841). Performance metrics for the final tuned models trained on the training set data are shown in online supplemental table 2. The SVM model had the highest accuracy (81.2%) and precision (0.855). The RF model had the best F₁-score (0.858), recall (0.918) and ROC-AUC (0.872) out of all models. The XGBoost model had the highest specificity (0.676). The MLP model had lower performance metrics than the other three models.

Table 2. Performance metrics after evaluation of each model on the test set data.

	RF	SVM	MLP	XGBoost
Accuracy	0.784	0.767	0.757	0.815*
F1-score	0.841	0.821	0.801	0.859*
Precision	0.798	0.808	0.841*	0.841*
Recall	0.888*	0.834	0.765	0.877
Specificity	0.600	0.648	0.743*	0.705
ROC-AUC	0.898*	0.845	0.826	0.882

^*Optimal values for each metric are emphasised with an asterisk.

MLP, multilayer perceptron; RF, random forest; ROC-AUC, area under the receiver operating characteristic curve; SVM, support vector machine; XGBOOST, extreme gradient boosted decision tree.

The RF model feature importance plot in figure 3A shows a gradual decrease in importance over the variables included in the study, indicating most features were important for prediction of disease activity. Features that contributed most to the RF model were neutrophils, platelets, haematocrit, infliximab and monocytes. The cut-off point for features retained for future training iterations was chosen at the vertical plateau, after faecal calprotectin. Features not retained for future training runs of the RF model were ESR and vedolizumab.

Figure 3. (A) Feature importance plot for random forest model. Features are ranked on the y-axis by their importance score. (B) Shapley additive explanations (SHAP) value analysis for XGBoost model. Negative SHAP values demonstrate feature values that decrease the probability of the predicted outcome (remission), whereas positive SHAP values demonstrate feature values that increase the probability of the predicted outcome (active disease). Points coloured in grey are missing values. ESR, erythrocyte sedimentation rate; XGBoost, extreme gradient boosted decision tree.

SHAP analysis results in figure 3B show the relative importance of the features used in the XGBoost model. CRP, albumin, neutrophils, haematocrit and infliximab were the most important variables in predicting disease activity. The global SHAP values (next to the variable names on the y-axis) indicate that the CRP is the most useful of our variables for classifying disease activity, with higher values increasing the probability of active disease. Other variables had substantially lower impacts on the model.

Patient-specific statistical analysis

Welch’s t-test was performed to identify significant differences in bloodwork value means between active disease and remission in individual patients (online supplemental table 3). This table lists the most significant metrics for each patient. Two patients had at least one bloodwork metric that was significantly different between active disease and remission. Patient 10 had a significantly higher haemoglobin in remission compared with active disease (t(23.63)=12.52, p_Bonf<0.001). Patient 8 had significantly higher lymphocytes in remission than in active disease (t(39.70)=4.56, p_Bonf=0.009). These trends were seen in other patients such as patient 5 (p_Bonf=0.104), patient 1 (p_Bonf=1.620) and patient 16 (pP_Bonf=0.938), but the results were not significant after multiple testing correction. Additional bloodwork metrics that had the largest differences between active disease and remission in individual patients, while not being significant, were platelets, neutrophils, monocytes, CRP, haematocrit and basophils.

There were two patients with significant differences in their bloodwork between active disease and remission (online supplemental figure 4). As patient 10 in the top graph achieved remission, their haemoglobin values increased to around 140 g/L, compared with previous values closer to 100 g/L. On the other hand, patient 8 in the bottom graph fluctuated between active disease and remission. It is still apparent that lymphocyte values in remission tend to be higher, around $2.5\times {10}^9$ cells/L compared with those in active disease, where the values range around $1.8\times {10}^9$ cells/L.

WBC differential and medications

The effect of medication is an important consideration in interpreting our results, since the majority of patients with a diagnosis of IBD are on medication. Multivariate analysis showing the initial MANOVA identifies acetylsalicylic acid (F(5, 1452)=6.12, p_Bonf<0.001), infliximab (F(5, 1452)=26.69, p_Bonf<0.001), adalimumab (F(5, 1452)=3.97, p_Bonf<0.05), prednisone (F(5, 1452)=67.23, p_Bonf<0.001), azathioprine (F(5, 1452)=51.98, p_Bonf<0.001) and budesonide (F(5, 1452)=4.40, p_Bonf<0.01) as medications that have a significant effect on the WBC populations in IBD patients (online supplemental table 4). Vedolizumab (p_Bonf=0.21), methotrexate (p_Bonf=0.19) and ustekinumab (p_Bonf=1.98) did not have a significant effect on WBC differentials.

The effects of medications on individual WBC types from the MANOVA that were significant are shown in online supplemental table 5. Most notably, prednisone and azathioprine had the most significant effect on the WBC differentials. Patients taking prednisone had significantly different neutrophils (F(1, 1456)=233.22, p_Bonf<0.001), eosinophils (F(1, 1456)=84.48, p_Bonf<0.001) and basophils (F(1, 1456)=23.12, p_Bonf<0.001) compared with patients not taking prednisone. Between patients taking azathioprine and those not, all WBC types were significantly different: neutrophils (F(1, 1456)=32.19, p_Bonf<0.001), lymphocytes (F(1, 1456)=177.45, p_Bonf<0.001), monocytes (F(1, 1456)=94.69, p_Bonf<0.001), eosinophils (F(1, 1456)=23.77, p_Bonf<0.001) and basophils (F(1, 1456)=49.97, p_Bonf<0.001).

Discussion

This study found that ML models were able to classify IBD activity based on routinely collected, non-invasive biomarkers (online supplemental figure 5). The XGBoost model was statistically most robust with an ROC-area under the curve of 0.882. Important biomarkers across multiple models for the classification of disease activity were neutrophils, CRP and albumin. Statistical analysis also demonstrated that any individual biomarker alone was poor at classifying disease activity state. This supports the need for multivariate data analysis, such as ML, to solve this problem. Finally, the medications that patients were taking at the time of bloodwork collection had small but significant effects on biomarker levels, which justifies inclusion of medications in our ML modelling.

Current common clinical methods of IBD activity monitoring include endoscopy, biopsy, intestinal ultrasound (IUS), magnetic resonance enterography (MRE), faecal calprotectin testing or CRP testing.⁶ A study using the simplified Magnetic Resonance Index of Activity to monitor disease using MRE was able to predict active terminal ileal disease in CD patients to a sensitivity of 0.83 and specificity of 0.41.⁷ In CD, IUS can detect active disease to a sensitivity of 0.85 and a specificity of 0.91.⁸ A meta-analysis investigated the ability of faecal calprotectin and CRP as markers for monitoring of disease activity in 2499 IBD patients. The pooled sensitivity and specificity of these two markers were 0.88 and 0.73 for faecal calprotectin, and 0.49 and 0.92 for CRP, compared with the gold standard of endoscopy.⁹ These other models have similar performance to the metrics in our study, although our study’s biomarkers are likely considerably more accessible, through decreased turnaround time and cost. Notably, bloodwork will be a more acceptable test than faecal calprotectin for some patients, since bloodwork does not require collection of one’s own stool.

A recently proposed biomarker for inflammatory conditions besides IBD is the neutrophil-to-lymphocyte ratio (NLR). The NLR has been used to assess the severity of sepsis, due to the increasing neutrophil mobilisation in response to inflammation throughout the body.¹⁰ The differential used in our analysis includes both the neutrophil and lymphocyte count but takes the analysis a step further by integrating these values with other commonly measured biomarkers for IBD. Our approach, while more computationally costly, minimises the effects of collinearity among these two predictors.

Overall, the potential clinical applications of the models generated in this study include providing a less invasive method of disease monitoring than endoscopy or cross-sectional imaging that is also not reliant on subjective symptom reporting data.

One potential limitation of this study is that disease classification was based on a subjective physician assessment (PGAS) at the time of bloodwork. However, as seen in our individual patient-based analysis, disease activity may not always be reflected in the bloodwork simultaneously, so there may be a lag in when each parameter changes. Some lymphocyte values indicated as remission seem to match other values from active disease, and vice versa. This lag in bloodwork changes also reveals that there may be residual long-lasting bloodwork abnormalities that persist despite symptom resolution, or that bloodwork values change before symptom onset as patients begin a flare of active disease. Validation models with larger time-series data collection should be used to address this in future studies.

This study demonstrated that ML analysis was able to classify IBD activity from bloodwork and medication parameters alone. Neutrophil count, albumin and CRP were consistently the most important metrics that changed with disease activity. Individual-by-individual analysis of these parameters did not demonstrate that there are substantial intra-individual differentiators of activity, but rather that the parameters show a complex statistical interplay, requiring a multivariate analysis approach. In the future, ML models could be applied to clinical practice to reduce the need for repeat colonoscopies and invasive procedures in IBD patients. This could also lessen the use of healthcare resources.

Data availability statement

No data are available.