Data collection process in strict implementation of special management, the full name of quality control. Special personnel are responsible for collecting clinical images and pathological information of patients in the medical record system, and personal identification data will be used to identify and process, in order to protect the privacy of test patients/participants. At the same time, there is a special person responsible for quality control and source data verification. Participants use the project unified number as the unique identification code, personal data and case information by the sample management personnel input and save, the user only see the individual number, no longer show the participants' name and other personal information. Data dictionary that contains detailed descriptions of each variable used by the registry, including the source of the variable, coding information was built. Participants' sample information is stored electronically in a dedicated computer that is not used for other purposes and is provided with a password, which is only available to the person (1 person) who manages the sample. Participants' medical records will be kept at the hospital and will be accessible only to researchers; If necessary, members of the bidding organization, ethics committee or government management department may access the personal data of the participants according to the corresponding authority. The results of the study will be published as statistically analyzed data and will not contain any identifiable participant information.
The sponsor is responsible for implementing and maintaining quality assurance and quality control systems with written Standard Operating Procedures (SOPs) to ensure that trials are conducted and data are generated, documented (recorded), and reported in compliance with the protocol, good clinical practices (GCP), and the applicable regulatory requirement(s). The SOPs should cover system setup, installation, and use. The SOPs should describe system validation and functionality testing, data collection and handling, system maintenance, system security measures, change control,data backup, recovery, contingency planning, and decommissioning. The responsibilities of the sponsor, investigator, and other parties with respect to the use of these computerized systems should be clear, and the users should be provided with training in use. Noncompliance with the protocol,SOPs,GCP, and/or applicable regulatory requirement(s) by an investigator/institution, or by member(s) of the sponsor's staff should lead to prompt action by the sponsor to secure compliance.
According to the morbidity of LARC and the risk of distant metastasis, the sample size is 300, and the follow-up interval is at least 3 years.
Plan for missing data: If the proportion of missing data is very large, such as greater than 95%, the investigators can directly remove this field; At 50\~95%, the investigators have two processing methods, one is to remove this field directly; another way is to turn the field into an indicator variable; that is the 0-1 variable. If the field is empty, the field is 0; Otherwise the field is 1. Between 5% and 50% : In this scenario, the investigators need to fill in the missing values. In the process of filling, there are two categories: simple filling and algorithm filling. Simple filling includes: 0 filling, mean filling, median filling, mode filling; Algorithm filling methods such as K Nearest Neighbors (KNN) filling, random forest filling and so on.
Statistics Statistical Product and Service Solutions (SPSS, version 26.0) and R software (version 4.0.5) were used for statistical analyses. Receiver operating characteristic (ROC) curve analysis was used to evaluate the optimum cutoff value of tumor stromal ratio (TSR) in discriminating DM risk based on the maximum Youden index. Heat maps showed the distribution of variables between patients with or without DM within 3 years. The independent DMFS risk factors were determined using Kaplan-Meier (K-M) curves and Cox regression analysis sequentially based on the data of the training cohort. Statistical significance was set at P\<0.05. Inter-observer variability was assessed using κ statistics for categorical and ranked variables, and ICC for continuous variables.
TSR assessment Biopsy specimens from colonoscopy were sectioned into 5 μm slices and stained with H\&E. Areas with both stromal and tumor cells presented on all four sides were selected to evaluate tumor stroma ratio (TSR) using an automated scoring method. The highest proportion of stromal components in all measured areas was recorded as the final TSR value in this study.
Magnetic Resonance analysis Machine-learning method was used to analyze the MR images, which includ image acquisition and reconstruction, image segmentation, feature extraction and qualification, analysis, and model building.
Pretreatment magnetic resonance (MR) examinations were conducted using 3.0 T scanners with an 8-channel phased-array wrap-around surface coil. An intramuscular injection of 10 mg raceanisodamine hydrochloride was administered to minimize bowel movement unless contraindicated. Sequences acquired included T1-weighted imaging (T1WI), T2-weighted imaging (T2WI) with and without fat saturation, and diffusion-weighted imaging (DWI) .
The tumor region of interest (ROI) was manually delineated slice-by-slice on high-resolution oblique axial T2WI (orthogonal to the rectal lumen) by the first radiologist and subsequently confirmed by the second radiologist with more experience on Insight Segmentation and Registration Toolkit-Standford Network Analysis Project (ITK-SNAP), from which the three-dimensional whole tumor volume of interest (VOI) was obtained. Disagreements were resolved through discussions. The radiologists were blinded to the clinicopathological information. Overall, 1229 features were extracted from each VOI, which can be classified into four categories: (1) shape characteristics; (2) first-order statistical characteristics; (3) texture features; and (4) high-order statistical characteristics.
The extracted features' inter- and intraclass correlation coefficients (ICCs) were calculated to assess the reproducibility of the features. Features with \<0.75 ICCs were considered non-stable and were eliminated. Pearson's correlation analysis was used to identify redundant features, and for any two features with a coefficient of 0.9, the one with the larger mean absolute coefficient was eliminated. The least absolute shrinkage and selection operator algorithm (LASSO) was applied to select the most significant predictive parameter from the training cohort, and 5-fold cross-validation was used to perform dimensionality reduction. A signature (i.e. Radscore) was calculated using a linear combination of the final selected features weighted by the respective coefficients.
Model built and validation According to the results of Cox regression, a nomogram (Mr) integrating all independent risk factors except TSR, a TSR nomogram (Mt) integrating all independent risk factors except the Radscore, and a combined model (Mrt) incorporating all the independent risk factors were constructed to predict the 3-year DM risk. The discriminative ability of these models was evaluated and compared using ROC curves. Calibration plots were drawn to explore the calibration ability of the three models. Decision curve analysis was performed to explore the clinical benefits by calculating the net benefit of each decision strategy at each threshold probability.
