Machine learning-based phenogroups and prediction model in patients with functional gastrointestinal disorders to reveal distinct disease subsets associated with gas production

Patients and data sources

This was a prospective cohort study. This study was approved by the ethics committee of Peking University Third Hospital (2021-132-01) and was conducted in accordance with the ethical guidelines outlined in the Declaration of Helsinki. Informed consent was obtained from the participants on an opt-out basis given the noninvasive nature of the study.

All patients with GI symptoms referred to the Outpatient Department of Gastroenterology at Peking University Third Hospital from December 2021 to November 2023 were eligible for inclusion in this study. FGIDs were established according to Rome IV criteria. The exclusion criteria were described in Supplementary Method 1. All medical history-taking was conducted by skilled gastroenterologists (Supplementary Method 2). Standardized questionnaires were used to classify GI symptoms into FGIDs according to the Rome IV criteria, and contained questionnaires assessing the severity of anxiety, depression, and sleep quality (Supplementary Table 1). The intensity of GI symptoms was assessed using a 4-point Likert scale ranging from 0 (none) to 3 (severe).

Breath tests

All patients underwent standardized lactulose hydrogen methane breath tests (LHMBT) and hydrogen sulfide breath tests performed by the same technician (Supplementary Method 3). Excessive gas production was defined as a positive LHMBT according to the North American Consensus, with an increase of hydrogen (H₂) ≥20 parts per million (ppm) or methane (CH₄) levels reaching ≥10 ppm within 90 min.^[16]

Unsupervised machine learning

Forty-three baseline variables, including clinical characteristics, H₂ peak, excessive H₂ production, CH₄ peak, excessive CH₄ production, hydrogen sulfide (H₂S) peak, time to H₂S peak, and the area under the curve (AUC) of H₂S, were identified and utilized as input for the unsupervised ML algorithm (Supplementary Method 4). The dataset was supplemented using the multiple imputation method to fill in missing data representing less than 40% of the total data. Twenty complete datasets were imputed. The variation analyses of the 20 imputation datasets were provided (Supplementary Table 2).

Individuals were grouped and different categories of FGIDs were identified based on their phenotypic characteristics using the K-means algorithm. Each imputed dataset was individually processed using the prescribed steps for the clustering algorithm.^[17] The clustering algorithm was executed with a range of predefined group numbers, starting from 2 and increasing to 9. The elbow method was used to determine the optimal number of clusters (Supplementary Figure 1). Principal components analysis (PCA) was used to visualize and investigate the clustering of our data (Supplementary Figure 2).

Supervised machine learning

Variables associated with excessive gas production were included as candidate variables based on a literature review, expert opinions, and the results of univariate analyses. Multicollinearity between variables was analyzed using Spearman correlation analysis and collinearity diagnostics (Supplementary Method 5). Least absolute shrinkage and selection operator (Lasso) regression was employed on all 20 imputed datasets to assess the importance of potential variables. To create a robust and dependable model, we chose the top seven characteristics for constructing the model based on a majority vote.

Internal validation was performed using a bootstrap method. Six classical ML models, namely logistic regression (LR), linear discriminant analysis (LDA), k-nearest neighbor (KNN), naive Bayes (NB), support vector machine (SVM), and random forest (RF), were developed to predict excessive gas production. The performance of the model was evaluated by examining the receiver operating characteristic (ROC) and calibration curves. Furthermore, we used decision curve analysis (DCA) to visually evaluate the net benefits of each model. The SHapley Additive exPlanations (SHAP) method was employed to visualize the importance of characteristics selected in the model. A dynamic nomogram was constructed for clinical application.

Statistical analysis

Categorical variables were presented as percentages and were analyzed using the Chi-squared test. Variables displaying a non-normal distribution were presented as medians along with their interquartile ranges and analyzed using the Mann-Whitney U-test. Multiple comparisons of breath gas features were analyzed using the Kruskal-Wallis test with Bonferroni post hoc correction, and repeated measurements of gas concentrations were analyzed using generalized estimating equation (GEE) models in R 4.3.0 (R Foundation for Statistical Computing, Vienna, Austria). The ML algorithm was executed using Python 3.11.5 (Python Software Foundation, Wilmington, USA) and Project Jupyter 6.5.4 (Anaconda, Inc., Austin, USA). Statistical significance was determined for all two-sided P values < 0.05.

Patient baseline characteristics categorized by phenogroups

FGIDs were confirmed in 300 of the 508 patients. A flowchart of the patients involved in this study was shown in Figure 1. Using the Rome IV criteria, functional dyspepsia (FD) was the most common FGID (45.0%), followed by irritable bowel syndrome (IBS; 42.7%, n = 128; Table 1, Supplementary Figure 3 and 4A). Overlapping of FGIDs were seen in 115 (38.3%) patients (Table 1).

Figure 1

Flow diagram of patient cohort. FGIDs: functional gastrointestinal disorders.

Excessive gas production was observed in 183 patients (61.0%, Table 1), with IBS representing the largest proportion (42.6%, n = 78), followed by FD (39.9%, n = 73; Supplementary Table 3 and Supplementary Figure 4B).

The K-means algorithm was used to integrate clinical symptoms and breath gas characteristics for phenogrouping. The most statistically significant clustering solution divided the entire patient population into four clusters using the elbow method based on a majority vote.

Cluster 1: Lower abdominal discomfort with the highest proportion of psychological burden

The cluster comprised 19.7% (n = 59) of the total participants and was characterized by higher-than-average scores for psychiatric scales (Table 1, all P < 0.05). Those patients presented more severe complaints of lower abdominal discomfort (P < 0.001) compared with the other three clusters, and many of them overlapped with other GI symptoms. The cluster had a higher proportion of extraintestinal symptoms such as oral ulcers. This cluster had the lowest proportion of excessive gas production (P < 0.001).

Cluster 2: Diarrhea predominance with a high prevalence of excessive gas production

This cluster consisted of 46.3% (n = 139) of the cohort (Table 1, Figure 2). The duration of this cluster was longer (P = 0.007). Most reported type 6/7 in the Bristol stool form (BSF). Many patients have reported experiencing abdominal pain and symptoms associated with gas production, such as borborygmus and increased flatus.

Figure 2

Description of clinical characteristics among clusters using a single spider plot superimposing the signature of each cluster. The values depicted were normalized to the entire population. Green, yellow, red, and deep purple indicate Clusters 1, 2, 3, and 4, respectively. MBT: methane breath tests, HBT: hydrogen methane breath tests, LHMBT: lactulose hydrogen methane breath tests, PSQI: Pittsburgh sleep quality index, HAMD: Hamilton Depression Scale, HAMA: Hamilton Anxiety Scale, BSF: Bristol stool form, PPI: proton pump inhibitor, BMI: body mass index.

This cluster had an above-average proportion of excessive gas production.

Cluster 3: Upper abdominal discomfort with a high prevalence of excessive gas production

This cluster comprised 21.3% (n = 64) of the total cohort (Table 1, Figure 2). The patients in this cluster were significantly younger, had a lower body mass index (BMI), and had a relatively short disease duration (all P < 0.05). This cluster was distinguished by a higher percentage of individuals experiencing upper abdominal discomfort (Supplementary Figure 5). This cluster had a relatively high proportion of excessive gas production, and most met the diagnosis of FD.

Cluster 4: Diarrhea predominance combined with postprandial symptoms presenting the highest prevalence of excessive gas production

This cluster comprised 12.7% (n = 38) of the cohort. Many of these patients had a history of allergic diseases. Similar to Cluster 2, patients in this cluster presented with loose stools and moderate lower GI symptoms. Many of them had postprandial symptoms. The cluster had the highest proportion of excessive gas production (100%, P < 0.001). The lowest scores were noted for the anxiety and depression scales (P = 0.001).

Gas time profiles after lactulose ingestion

The gas time profiles of the different clusters were shown in Figure 3. Temporal trends and variations in breath gas following lactulose ingestion in different clusters of patients with FGIDs were analyzed using a generalized estimation equation (Supplementary Table 4 and 5). The levels of exhaled H₂, CH₄, and H₂S showed significant variations following ingestion of lactulose (all P < 0.001, Supplementary Table 4).

Figure 3

Gas time profiles in patients with FGIDs (n = 300). (A) Breath hydrogen concentration-time profiles grouped by clusters following lactulose ingestion. (B) Breath methane concentration-time profiles grouped by clusters following lactulose ingestion. (C) Breath hydrogen sulfide concentration-time profiles grouped by clusters following lactulose ingestion. (D) Gas time profiles in Cluster 1. (E) Gas time profiles in Cluster 2. (F) Gas time profiles in different Cluster 3. (G) Gas time profiles in Cluster 4. Data were mean ± SD. The differences of gas production between these clusters were examined by GEEs with time, group (Cluster 1, Cluster 2, Cluster 3 and Cluster 4), and group-by-time interaction as the covariates. Pairwise comparisons were performed with Bonferroni corrections. ^**Cluster 1 vs. Cluster 4, P < 0.01; ^***Cluster 1 vs. Cluster 4, P < 0.001; ^{^}Cluster 2 vs. Cluster 4, P < 0.05; ^{^^}Cluster 2 vs. Cluster 4, P < 0.01; ^{^^^}Cluster 2 vs. Cluster 4, P < 0.001; ^#Cluster 3 vs. Cluster 4, P < 0.05; ^###Cluster 3 vs. Cluster 4, P < 0.001. FGIDs: functional gastrointestinal disorders, ppm: parts per million, ppb: parts per billion, H₂: hydrogen, CH₄: methane, H₂S: hydrogen sulfide, SD: standard deviation.

The increasing trends of exhaled H₂ concentrations were similar among the four clusters (Figure 3A). However, significant differences were observed in both the H₂ peak and H₂ AUC among the four clusters (all P < 0.001, Table 2). Compared to Cluster 1 and Cluster 3, Cluster 2 and Cluster 4 had higher H₂ peaks and AUCs, especially in Cluster 4 (all P < 0.001, Table 2). Further analysis using GEEs showed that H₂ concentrations were significantly elevated at multiple time points in Cluster 4 (all P < 0.001, Supplementary Table 5).

The temporal trends of exhaled CH₄ concentration were relatively flat among Cluster 1, 2, and 3, whereas CH₄ concentration showed a significant elevation at 75–90-min time points in Cluster 4 (P < 0.001, Figure 3B, Table2). Similar results were obtained using GEEs (Supplementary Table 5).

Re-increasing trends in H₂S were observed in all four clusters after a short period of decrease (Figure 3C). The change in H₂S concentration over time was distinct among the four clusters. Both multiple comparisons among clusters and the results of GEEs found that the H₂S concentration was significantly higher in Cluster 4 vs. the other clusters at multiple time points (P < 0.05, Supplementary Table 5). Notably, the total concentration of H₂S was not high in Cluster 3, but H₂S peaked earlier (P = 0.021, Table 2).

A supervised machine learning approach to predict excessive gas production

The FGIDs cohort was split into two sets: training with 300 patients and internal validation using bootstrap method. In the training dataset, 183 patients (61%) were identified as having excessive gas production, as defined by a positive LHMBT. Detailed clinical characteristics were provided in Supplementary Table 3.

Features selection

Based on literature reviews, expert opinions, and univariate comparisons between positive and negative LHMBT results in the training dataset, 36 features related to excessive gas production were identified. After adjusting for multicollinearity between variables, LASSO regression was utilized to pinpoint seven features for model construction: sex, BMI grade, drinking history, early satiety, fullness after meals, borborygmus, and abdominal distension. The predictors were utilized in the following modeling procedures (Supplementary Figure 6 and Supplementary Table 6).

Performance of the ML models in predicting excessive gas production

We conducted an analysis of the predictive performance of various ML models, as illustrated in Figure 4. The models exhibited varied capabilities for predicting excessive gas production, as defined by the positive LHMBT (Figure 4A). Among the six classification algorithms, KNN demonstrated superior performance, achieving an AUC of 0.85 for the cohort, demonstrating good discrimination ability (Figure 4A). The ML models exhibited robust performance in discriminating excessive gas production within the internal validation cohort, as evidenced by an AUC ranging from 0.71 to 0.74 (Figure 4B).

Figure 4

Performance evaluation of ML models for predicting excessive gas production in FGIDs. (A) ROC curves of different models adopted for predicting excessive gas production in the FGIDs cohort training dataset. (B) ROC curves of different models adopted for predicting excessive gas production in the FGIDs cohort internal validation dataset. (C) Calibration plot of different models for predicting excessive gas production in patients with FGIDs. The model-predicted probability of excessive gas production was plotted on the x-axis and the actual proportion of excessive gas production was plotted on the y-axis. An ideal calibration plot is indicated by a 45-diagonal line. (D) Decision curve analysis of the model. The x-axis indicates the threshold probability of the excessive gas production. The threshold probability is the level of certainty above which a patient or physician would choose to intervene. The y-axis indicates the net benefit, which is calculated as true positive rate – (false positive rate × weighting factor). ML: machine learning, FGIDs: functional gastrointestinal disorders, ROC: receiver operating characteristic curve, LR: logistic regression, LDA: linear discriminant analysis, KNN: k-nearest neighbor, NB: naive Bayes, SVM: support vector machine, RF: Random Forest, AUC: area under the curve, CI: confidence interval.

Furthermore, the calibration plots verified the agreement between the predicted excessive gas production from the ML models and actual observations (Figure 4C). Additionally, DCA revealed that by setting a probability threshold for excessive gas production occurrence within the range of 38.1% and 79.9%, utilizing the model yielded significantly superior net benefits compared to both strategies of “treating none” and “treating all” (Figure 4D). These findings highlighted substantial clinical applicability of our prediction model.

Interpreting models using the SHAP approach

We conducted a comprehensive investigation of the fundamental workings of our ML model using the SHAP method (Figure 5). The SHAP summary plot revealed that female individuals with a history of drinking, low BMI, moderate abdominal distention and postprandial symptoms, and exaggerated bowel sounds were more likely to be identified as having excessive gas production.

Figure 5

SHAP summary plot for the predictors contributing to the ML model prediction of excessive gas production. The organization of variables in the plot showcased their significance, with color coding showing the size and direction of impacts. A higher likelihood of having excessive gas production was represented by SHAP values exhibiting positive magnitudes. SHAP: SHapley Additive explanation, ML: machine learning.

Finally, we successfully developed a user-friendly and easily accessible web-based nomogram for predicting excessive gas production (https://jxl-xs.shinyapps.io/DynNomapp/).