
Abstract Background Breast cancer is the most prevalent cancer in women, and postoperative disease-free survival (DFS) is an important indicator of long-term prognosis. However, tumor, node, metastasis staging has limited ability to integrate multidimensional clinical, pathological, and treatment-related factors for individualized DFS prediction. Therefore, machine learning (ML) may provide a more comprehensive approach to prognostic risk assessment. Objective This study aimed to develop and externally validate a ML-based model for predicting postoperative DFS in patients with breast cancer. Methods This retrospective study included 2 cohorts: 1231 patients treated at an institutional center between 2012 and 2020 and 304 patients from The Cancer Genome Atlas (TCGA). The institutional cohort was divided into model development (n=861) and internal validation (n=370) sets, while the TCGA cohort served as an external validation cohort. Five ML algorithms were evaluated, with the Cox proportional hazards model additionally fitted as a conventional statistical reference. Model performance was assessed using the concordance index and time-dependent area under the curve (AUC). The net reclassification index and integrated discrimination improvement were used for comparative evaluation. Shapley additive explanations (SHAP) analysis was performed to assess predictor contributions, and patients were stratified into risk groups according to the median predicted risk score. Results Based on a comprehensive evaluation of model performance, extreme gradient boosting (XGBoost) with an accelerated failure time (AFT) was selected as the final model. The study included 1231 institutional patients, with 861 used for model development and 370 for internal validation, and 304 patients from TCGA for external validation. The 5-year time-dependent AUCs were 0.840, 0.822, and 0.762 in the training, internal validation, and external validation cohorts, respectively. In the external validation cohort, the XGBoost (AFT) model achieved a concordance index of 0.750 (95% CI 0.654‐0.843), with 3-, 5-, and 8-year AUCs of 0.791, 0.762, and 0.768, respectively. At 5 years, compared with Cox regression, XGBoost (AFT) showed improved reclassification (net reclassification index=0.311, 95% CI 0.172‐0.458; P =.01; integrated discrimination improvement=0.166, 95% CI 0.095‐0.235; P <.001). SHAP analysis identified endocrine therapy, pT, pN, lymph node ratio, Ki-67, and chemotherapy among the major contributors to model predictions. Risk stratification based on the median predicted risk score separated patients into groups with distinct DFS outcomes (hazard ratio 5.44, 95% CI 4.04‐7.31). Conclusions The ML-based model demonstrated favorable performance and external generalizability in predicting postoperative DFS. SHAP analysis provided interpretability for the model, and risk stratification identified patient subgroups with distinct prognostic risks.
Background Major above-ankle amputation is a devastating outcome of diabetes-related foot disease. The Wagner classification is routinely available but places many inpatients in the same grade; few prediction models for major amputation have undergone temporal validation or reported calibration. Objective This study aimed to develop an admission-time model for in-hospital major amputation from routinely collected data and to validate it internally in a later period at the same center. Its intended role is to support prognostic communication and prioritization of early multidisciplinary and vascular assessment and not to inform the decision to amputate. Secondary objectives were to compare machine learning algorithms with penalized logistic regression and to assess the added value of segmental limb pressures. Methods In this retrospective cohort study at a tertiary center in Taiwan, adults with a single hospitalization for diabetes-related foot disease from 2015 through 2025 were divided by admission date into a development cohort (2015-2021) and a temporal test cohort (2022-2025). Twenty admission variables were modeled. Five algorithms were compared by nested cross-validation in the development cohort; the one with the lowest out-of-fold log loss became the primary model, with operating thresholds fixed before a single evaluation in the test cohort. Results The development cohort included 1400 patients with 126 (9.0%) major amputations and the test cohort 771 patients with 51 (6.6%). In the test cohort, the primary ridge logistic regression model had an area under the receiver operating characteristic curve (AUROC) of 0.882 (95% CI 0.841-0.917) vs 0.804 (95% CI 0.737-0.867) for Wagner grade alone. Absolute risk was overpredicted: mean predicted risk was 9.6% vs 6.6% observed, with calibration intercept –0.47, slope 1.33, and observed-to-expected ratio 0.69 (95% CI 0.53-0.86). At the prespecified screening threshold, 343 (44.5%) patients were flagged; sensitivity was 0.980 (95% CI 0.897-0.997), specificity 0.593 (95% CI 0.557-0.628), positive predictive value 14.6% (95% CI 11.2%-18.7%), and negative predictive value 0.998 (95% CI 0.987-1.000). Of 771 patients, no major amputation occurred among the 396 (51.4%) patients in the lowest risk band; of 54 patients, 22 (40.7%) patients in the highest band underwent major amputation. Random forest and gradient boosting reached higher AUROCs (0.915 and 0.912) in secondary comparisons. Segmental limb pressures did not clearly improve discrimination in complete cases (change in AUROC –0.004). Conclusions In an internal temporal validation at a single center, a transparent model based on routine admission data added discrimination for in-hospital major amputation beyond Wagner grade alone but overpredicted absolute risk. It estimates amputation under 1 institution’s care pathway and was not compared with multidimensional classifications such as Wound, Ischemia, and foot Infection; whether acting on it improves outcomes is untested. Independent external validation, recalibration where needed, and prospective evaluation are required before clinical use; the model must not be used to indicate amputation or withhold limb-salvage evaluation.
Abstract Background The Sequential Organ Failure Assessment (SOFA) score is central to risk stratification in intensive care units (ICUs), but its thresholds and 0- to 4-point increments were set by expert consensus in 1996 and never reestimated from outcome data. Machine learning models discriminate better while discarding the additive, point-based structure that makes SOFA auditable at the bedside, and methods for updating a legacy score rather than replacing it remain scarce. Objective This study aimed to develop and internally validate a data-driven recalibration of a SOFA-derived score that learns thresholds and integer weights from outcome data while preserving an additive, interpretable point-based structure and to test whether it improves discrimination for ICU mortality over a prespecified SOFA-derived reference score. Methods We conducted a retrospective multicenter cohort study using the eICU Collaborative Research Database (208 US hospitals, 2014‐2015). We analyzed 14,848 adult first ICU admissions with complete day 1 SOFA-related measurements; 133,684 (90%) of 148,532 encounters with outcome data were excluded because of at least 1 missing determination. As vasoactive dose units could not be harmonized and urine output was largely unavailable, a prespecified SOFA-derived reference score served as the comparator. The framework couples supervised discretization (univariate decision trees and interval number chosen by a penalized validation criterion) with constrained integer-weight optimization under monotonicity, integrality, and scale-anchoring constraints, plus logistic calibration. Performance was estimated from pooled out-of-fold predictions of 10-fold nested cross-validation grouped by hospital, with bootstrap CIs; final thresholds and weights were refit on the full cohort and contributed no performance estimate. Results ICU mortality was 17.03% (2529/14,848). The recalibrated score achieved an area under the receiver operating characteristic curve (AUROC) of 0.811 (95% CI 0.800‐0.819) versus 0.751 (95% CI 0.740‐0.761) for the reference score (paired difference 0.059, 95% CI 0.053‐0.067), and an area under the precision-recall curve of 0.495 (95% CI 0.477‐0.515) versus 0.393 (95% CI 0.377‐0.411; paired difference 0.102, 95% CI 0.087‐0.116). Improvement was consistent across all 10 hospital-grouped folds and length-of-stay strata. After unweighted logistic recalibration, global calibration was adequate; the Brier score was lower (0.112 vs 0.123), whereas other calibration-error measures were higher. Discrimination approached unconstrained XGBoost (Extreme Gradient Boosting; AUROC 0.818) and exceeded penalized logistic regression (AUROC 0.775) fitted on the same folds. The advantage persisted in all 1000 resampling replications and in a prespecified sensitivity analysis expanding the cohort to 61,469 encounters (median paired AUROC difference 0.037). Conclusions Data-driven recalibration under explicit structural constraints can update the cutpoints and integer weights of an established clinical score without abandoning its additive, bedside-computable form and should transfer to other legacy instruments. The recalibrated score is a mortality risk model on its own 0‐71 scale; an open-source calculator is provided, and external and prospective validation is required before clinical use.
Abstract Background Peripheral blood smear (PBS) examination provides morphological information on immune cells that reflects systemic pathophysiology, yet whether it can be used to predict systemic states has not been systematically established. Objective This study aimed to develop and evaluate an integrated deep learning framework that leverages cell-level morphology from routine PBS examinations to discriminate acute inflammatory states and chronic comorbidities. Methods An integrated deep learning framework was developed for PBS analysis, in which cell-level morphological features were extracted and aggregated into slide-level representations to predict systemic states. A total of 51,369 PBS examinations were used for self-supervised cell-level pretraining and 9211 for slide-level downstream tasks. Outcomes comprised acute inflammatory states and chronic comorbidities. Acute inflammatory states included white blood cell (WBC) count (categorized as low, normal, or high), elevated C-reactive protein (CRP), elevated erythrocyte sedimentation rate (ESR), and septic shock. Chronic comorbidities included the Charlson Comorbidity Index (CCI), cardiovascular disease (CVD), chronic kidney disease (CKD), and diabetes mellitus (DM). Discriminative performance was assessed using the area under the receiver operating characteristic curve (AUROC) with 95% CIs, computed on the test set of the development dataset and on the holdout dataset. The framework was compared with a demographic model based on age and sex and with a cell-type-only model without morphological features. As a secondary analysis, temporal validation was performed using a chronological split. Results The framework discriminated acute inflammatory states, achieving development and holdout AUROCs of 0.852 (95% CI 0.836‐0.867) and 0.849 (0.833‐0.864) for WBC count, 0.891 (0.870‐0.913) and 0.896 (0.876‐0.915) for CRP, 0.838 (0.794‐0.878) and 0.844 (0.799‐0.882) for ESR, and 0.849 (0.807‐0.886) and 0.822 (0.783‐0.853) for septic shock. Discrimination of chronic comorbidities was more modest but consistent across the 2 datasets, with AUROCs of 0.792 (0.762‐0.818) and 0.781 (0.751‐0.808) for CCI, 0.696 (0.665‐0.727) and 0.680 (0.654‐0.708) for CVD, 0.724 (0.690‐0.755) and 0.728 (0.696‐0.756) for CKD, and 0.645 (0.610‐0.680) and 0.638 (0.603‐0.674) for DM. The framework significantly outperformed the cell-type-only model for all outcomes in both datasets but did not significantly outperform the demographic model for CVD in the holdout dataset or for DM in either dataset. AUROCs in the temporal holdout dataset ranged from 0.621 for DM to 0.883 for CRP, consistent with those in the temporal development test set (0.666 to 0.889). Conclusions An integrated deep learning framework for automated PBS analysis can provide information associated with inflammation-related laboratory results and clinical outcomes, with stronger discrimination for acute inflammatory states than for chronic comorbidities. These exploratory findings suggest the potential of deep learning-based PBS analysis as a novel approach to assessing systemic states in clinical practice.
Abstract This article provides a critical analysis of system-system interoperability (SSI) in the health data domain. By SSI, we mean the idea of connecting technical systems with each other to exchange operational data, directly or via intermediate platforms or infrastructures. We start by giving a brief history of some uses of the word “interoperability,” from the Institute of Electrical and Electronics Engineers (IEEE) original definition in 1991 to the Interoperable Europe Act of 2024. We thereby identify ambiguity and hidden assumptions in the definitions themselves. We then establish precise terminology to make a stringent analysis possible. Aided by a simple case of SSI from our own clinical practice and using a method of logical analysis, we look closely at (1) the assumption that the semantics in SSI can be handled by referring to external semantic standards, (2) the assumption that SSI is a feasible route to the safe exchange of operational data between systems in primary use, and (3) the challenges encountered in scaling SSI solutions to involve multidirectional data exchange, multiple data types, and multiple systems. Our first conclusion is that semantic interoperability, as defined by the Healthcare Information and Management Systems Society (HIMSS) in 2019, may not be a feasible route to stabilizing the real-world meaning of data exchanged in SSI solutions. The main reason is that referring to external standards cannot capture enough contextual information to correctly interpret health data remote from the primary context in which they were first registered. Our second conclusion is that SSI may not be a feasible route to the exchange of operational data between systems in primary use. Multiple unresolved problems are identified—problems that cannot be overcome by adding new components to existing systems, as proposed in, for example, the European Health Data Space (EHDS) legislation. Instead, a thorough redesign of each system’s internal architecture, data model, and user interface to align them with each other would be required. It is far from clear how such redesign would be practically achievable. Our third conclusion is that SSI may not be a scalable solution. In part, this is a consequence of the difficulties in using external standards to handle the semantics. Instead, careful negotiations and agreements about the exact meaning of each data type to be exchanged would be needed to guard against misinterpretation and, hence, risks in both primary and secondary use. We conclude that interoperability should be repositioned from being a taken-for-granted pathway to “better use of health data” to being an explicit object of critical inquiry. Further research should establish a commonly agreed analytical framework for evaluating suggested solutions. This framework must include stringent definitions of concepts and terminology, theoretical requirements, and practical aspects to be addressed in a concrete implementation.