We analyzed anonymized, deidentified data from a multicenter cohort comprising individuals who underwent liposuction in 2024 across 20 obesity specialty clinics in the 365mc network in South Korea [10]. A comprehensive dataset was constructed from patients who underwent procedures at these clinics throughout 2024, totaling 8064 individuals. To ensure data integrity for model development, patients with incomplete records regarding the liposuction site or postoperative outcomes were excluded. After data preprocessing, the final cohort included 7804 patients. A detailed flowchart of the data processing is shown in Figure S1 in Multimedia Appendix 1.

The research protocol was approved by the institutional review board of Kyung Hee University (KHUH 2024–04–002). The requirement for informed consent was waived by the board, as deidentified data were used for all analyses. This study was conducted in accordance with the principles outlined in the Declaration of Helsinki. The Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis statement was followed to guide the design, analysis, and reporting of this study [11].

The multicenter liposuction cohort included individuals aged ≥17 years who underwent liposuction between 1 January and 31 December 2024 at 20 obesity specialty clinics in the 365mc network in South Korea. Clinical records from these branches were consolidated into a single dataset and randomly partitioned into training, validation, and test subsets for model development. The ML model was developed to simultaneously predict 2 primary postoperative outcomes: body weight and site-specific circumference, using preoperative variables [12]. Participants’ weight and the dimensions of the treated site were measured preoperatively and again at a mean of 3.82 (SD 2.24) weeks after liposuction with a bioelectrical-impedance analyzer; all measurements were obtained with the InBody 370S (InBody Co) [13]. Only patients with complete paired preoperative and postoperative anthropometric and bioelectrical data were included in the final analysis.

To develop a multioutput ML-based model for predicting postoperative outcomes after liposuction, we used the following 15 features: sex (male or female), age (continuous), height (continuous), preoperative body weight (continuous), BMI (continuous), preoperative size (continuous), liposuction technique (local anesthetic minimally invasive liposuction [LAMS] or conventional surgical liposuction), liposuction site (abdomen, arms, back, buttocks, calves, flanks, or thighs), skeletal muscle mass (continuous), body fat mass (continuous), total body water (continuous), fat-free mass (continuous), body protein (continuous), body mineral content (continuous), and waist-to-hip ratio (continuous). Detailed descriptions of the variables are provided in Table S2 in Multimedia Appendix 1.

The dataset was randomly split into training and test sets in a 7:3 ratio. The test set was used only for final model evaluation and was not accessed during training or internal validation. Missing values were handled using multiple imputation by chained equations. Continuous variables were winsorized to reduce the influence of outliers [14], and z-score standardization was then performed using the mean and SD estimated from the training set, which were subsequently applied to the test set [15]. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regularization, with detailed procedures provided in Multimedia Appendix 1 and the selected features shown in Figure S2 in Multimedia Appendix 1. The correlation matrix of the resulting feature set is shown in Figure S3 in Multimedia Appendix 1 [16,17].

We established a multioutput ML framework to predict postoperative body weight and circumferential size using preoperative profiles from the 365mc network. The candidate algorithms included tree-based ensemble models, including the random forest regressor and extra trees regressor; boosting-based models, including gradient boosting regressor (GBR), histogram-based GBR (HGBR), and adaptive boosting (AdaBoost) regressor; and a kernel-based model, support vector regressor [18,19]. The Tree-Based Pipeline Optimization Tool (TPOT; Epistasis Lab) was also evaluated as an automated ML approach [20]. Ensemble strategies, including the voting regressor and stacking regressor, were applied to combine predictions from multiple base learners [21]. For multioutput prediction, a regressor chain framework was implemented to model the dependency between postoperative body weight and circumferential size. The principles and key formulas of the algorithms used in this study are provided in Table S3 in Multimedia Appendix 1. To characterize the variability and distribution of the target variables, descriptive statistics were calculated separately for the training and test sets and are presented in Table S4 in Multimedia Appendix 1. All models were optimized using grid search [22], and the final hyperparameters are summarized in Table S5 in Multimedia Appendix 1.

Multioutput learning was used to jointly predict 2 related postoperative outcomes, postoperative body weight and circumferential size, from the same selected input features [23,24]; let x denote the predictor vector for a given patient, and let y=(y1, y2) denote the target outcomes. We implemented a wrapper-based regressor chain using the extra trees regressor as the base estimator, allowing the prediction of the first target to inform the second target [8]. In the forward chain, postoperative body weight was predicted first and then used as an additional feature to estimate circumferential size:

In the reverse chain, the order of the 2 targets was swapped. Both chain directions were evaluated to assess the influence of target ordering on model performance. The overall objective was to minimize the joint prediction error across both outcomes:

The extra trees regressor with the forward chain order was selected as the final multioutput prediction model based on its overall predictive performance.

Model performance was evaluated using 5-fold cross-validation on the training set, as described in Multimedia Appendix 1. Predictive performance was assessed using the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), with their definitions and formulas summarized in Table 1 [25]. Higher R² values and lower RMSE, MAE, and MAPE values indicated better performance. Model selection prioritized R², with RMSE used as a secondary criterion among models with comparable R² values; MAE and MAPE were additionally considered to provide a comprehensive assessment of predictive performance. Model interpretability was assessed using Shapley additive explanations (SHAP) values [26], and an ablation analysis was performed by sequentially removing the most influential predictors and evaluating the resulting changes in performance [27]. All analyses were conducted in Python (version 3.10.17; Python Software Foundation), NumPy (version 1.26.4; NumFOCUS), Pandas (version 1.5.3; NumFOCUS), Matplotlib (version 3.10.1; NumFOCUS), and Scikit-learn (version 1.6.1; NumFOCUS) [28].

To evaluate the reliability of the model’s continuous predictions, agreement with the observed postoperative measurements was assessed using the Bland-Altman methodology [29]. For each patient, the difference between the predicted and observed values was plotted against their mean, and the overall mean difference together with the 95% limits of agreement (LoA) based on the SD was derived. Separate analyses were performed for postoperative body weight and circumferential size. To assess model robustness across different patient populations, subgroup analyses were performed according to sex and age group (Table S7 in Multimedia Appendix 1). To facilitate clinical adoption, an AI-driven, web-based CDSS was subsequently developed, enabling surgeons to input preoperative data and receive real-time predictions of postoperative body weight and circumferential size outcomes at the point of care [30].

The multicenter 365mc cohort comprised 8064 patients who underwent liposuction between January 1, 2024, and December 31, 2024, at 20 obesity specialty hospitals in the 365mc network in South Korea. After excluding 3.2% (260/8064) incomplete records, 96.8% (7804/8064) eligible participants were included in the final analysis, as shown in Figure S1 in Multimedia Appendix 1. The excluded records included 3.2% (249/7804) records with missing postoperative body weight or treated-site circumference, 0.1% (8/7804) records with missing surgical-site information, and 0.04% (3/7804) records with missing sex information. Baseline characteristics are summarized in Table 2. The cohort included 7612 (97.54%) female participants and 192 (2.46%) male participants. The mean age was 35.3 (SD 8.6) years, and the mean BMI was 24.0 (SD 3.9) kg/m².

Table 3 summarizes the performance of all candidate ML models and approaches, evaluated using R², RMSE, MAE, and MAPE. Using 15 predictors, we compared 8 individual ML models, an automated ML approach, 2 ensemble approaches, and chained multioutput regression models for predicting postoperative body weight and circumferential size. In the 5-fold cross-validation of the training dataset, the chained extra trees regressor with a forward prediction order showed the best overall performance across all evaluation metrics, with an R² of 0.986 (95% CI 0.983-0.988), an RMSE of 1.916 (95% CI 1.680-2.151), an MAE of 1.042 (95% CI 1.006-1.078), and a MAPE of 1.839 (95% CI 1.765-1.913). The stacked model using the extra trees regressor and random forest regressor showed comparable performance, with an R² of 0.985 (95% CI 0.983-0.988), an RMSE of 1.932 (95% CI 1.684-2.179), an MAE of 1.052 (95% CI 1.020-1.085), and a MAPE of 1.859 (95% CI 1.763-1.925).

In the independent test dataset, the chained extra trees regressor with a forward prediction order achieved an R² of 0.980, an RMSE of 2.356, an MAE of 1.242, and a MAPE of 2.199. Among the models achieving the highest R² of 0.980, this model had the lowest RMSE and was therefore selected as the final model based on the model-selection criteria. HGBR showed comparable performance, with an R² of 0.980, an RMSE of 2.359, and an MAE of 1.110. The architecture of the selected chained multioutput regression model is illustrated in Figures 1 and 2. To further evaluate predictive reliability, Bland-Altman analyses stratified by surgical site are presented in Figure 3.

Figure 4 displays SHAP summary plots for the chained multioutput model. Figure 4A illustrates the first regressor, which predicts postoperative body weight, confirming that preoperative weight and related clinical variables exert the greatest influence on the output. Figure 4B depicts the second regressor, which incorporates the predicted postoperative body weight from the first stage as an additional feature when estimating postoperative circumferential size. In this second stage, the predicted weight becomes an important predictor alongside preoperative circumferential size, liposuction site, and surgical technique.

To quantify the marginal contribution of each key feature, we performed an ablation analysis by sequentially removing 4 influential features: preoperative size, liposuction site, liposuction technique, and BMI. Model performance was then reassessed after each feature removal. The exclusion of preoperative size resulted in the largest performance decrease, with R² decreasing from 0.980 to 0.975 and error metrics increasing from 2.356 to 2.868 for RMSE, from 1.242 to 1.419 for MAE, and from 2.199% to 2.445% for MAPE. Despite this decrease, the model maintained strong predictive performance, suggesting that the overall prediction framework was robust to the removal of individual features. Full results of the ablation analysis are provided in Table S6 in Multimedia Appendix 1.

To facilitate clinical implementation, we deployed the chained multioutput model as a web-based CDSS accessible (365mc AI platform [31]). The platform guides the user through 3 sequential input screens that capture demographic details, operative variables, and body composition measurements, after which it instantly returns personalized forecasts of postoperative body weight and circumferential size. Figure 5 illustrates the interface, showing the stepwise data-entry process and the final results page. All information entered by the user is transmitted over an encrypted connection and is automatically deleted once the prediction is generated, ensuring that no personally identifying data are retained or stored on the server.

This study developed and validated a chained multioutput ML model that jointly predicts postoperative body weight and circumferential size using 15 predictors from a multicenter liposuction cohort across 20 obesity specialty clinics in the 365mc network. Among the benchmarked single-output, automated ML, ensemble, and chained approaches, the chained extra trees regressor achieved the best performance (R²=0.980, RMSE=2.356), supporting the value of explicitly modeling the dependency between these two clinically related outcomes. In this structure, postoperative body weight was predicted first and then used to estimate postoperative circumferential size, underscoring the conditional relationship between the two outcomes. Feature-importance analysis showed that preoperative weight was the main predictor of postoperative body weight, whereas preoperative size, liposuction-related factors, and the predicted weight estimate most strongly influenced postoperative circumferential size. The model was implemented as a web-based CDSS, supporting its potential for preoperative planning and individualized patient counseling.

This study proposed a chained multitarget regression model to jointly predict postoperative body weight and circumferential size after liposuction. These two outcomes are clinically related, and predicting them separately may fail to capture the dependency between overall weight change and local body-contour change [10,32]. Unlike previous single-target approaches, our model used the prediction of one outcome as additional information for predicting the other. This structure allowed the model to reflect the conditional relationship between postoperative body weight and circumferential size, which is important for surgical planning, objective outcome assessment, and patient counseling [33]. SHAP analysis further supported this relationship by showing that the predicted postoperative body weight contributed meaningfully to the prediction of postoperative circumferential size [34].

From a methodological perspective, direct single-output regression models and chained regression models have different strengths. Direct models are simple and useful when target outcomes are relatively independent, but they may not capture dependencies between clinically related outcomes [7]. In contrast, chained regression models use the prediction of one target as additional information for another, allowing them to reflect intertarget relationships [7,8]. In this study, postoperative body weight and circumferential size were considered clinically related outcomes after liposuction, and the chained extra trees regressor achieved the best overall performance, suggesting that modeling this dependency was appropriate for our prediction task. However, because predictions in subsequent steps depend on those made in the preceding steps, careful validation of the prediction order and model robustness is required [35].

Ensemble approaches also showed strong performance, with voting providing stable averaged predictions and stacking allowing a meta-model to combine base learners more flexibly [36,37]. However, more complex models, including stacking and chained regression, require careful validation because their performance may depend on the dataset structure, target relationships, and prediction order. Therefore, although our findings support the usefulness of the chained multioutput approach, future studies should further evaluate different chaining orders and validate the model in external datasets to confirm its generalizability and robustness.

Building on these methodological and ML modeling advances, we deployed our proposed chained multioutput model as a web-based CDSS that simultaneously forecasts postoperative body weight and circumferential size while leveraging individual patient profiles. By presenting key features including the surgical site, technique, and conditional relationship between outcomes, the system may enhance personalized preoperative counseling and surgical planning [38]. Explicit modeling and communication of the interdependence between outcomes further bolster interpretability and may increase clinician and patient trust, underscoring the need for prospective evaluation of how these explanatory elements affect decision quality, patient satisfaction, and adherence to recommended care [39].

Despite promising performance, this study has several limitations. First, the proposed model was developed using a retrospective cohort from a single high-volume liposuction network in South Korea, and the study population was predominantly young Asian women with a relatively low BMI [40]. These characteristics may restrict the external validity of the findings when applied to other ethnicities, anatomical surgical sites, or clinical settings. Second, reliance on existing medical records introduces potential selection bias and missing data; although records with critical omissions were excluded, residual confounding and incomplete capture of relevant variables cannot be excluded [41]. Third, although the model effectively leverages the conditional dependence between postoperative circumferential size and body weight, it does not explicitly characterize higher-order interactions among other preoperative factors such as age, sex, and body composition, which could further enhance predictive performance. Finally, the surgical-site information was based on a single, operator-designated representative region, and definitions or granularity of those regions may vary across institutions, limiting the generalizability of findings related to procedural localization. Prospective validation in multicenter and more diverse populations, including assessments across finer anatomical subregions and evaluation of the stability of key predictive features, is needed to confirm and extend these results.

This study developed and validated a chained multioutput regression model to predict postoperative body weight and circumferential size using 15 predictors from a multicenter liposuction cohort across 20 obesity specialty clinics in the 365mc network. The model accounted for the dependency between the two postoperative outcomes and was implemented as a web-based CDSS. The key findings are as follows: