Deep Learning Model Predicts 5-Year Mortality in NSCLC

Per a JMIR Preprint authored by Jong Hyuk Lee, Ho Cheol Kim, Kyu-Won Jung, and Chang Min Choi, the team developed and validated deep learning models to predict 5-year mortality in non-small-cell lung cancer (NSCLC) using data from the Korea Central Cancer Registry (KCCR). The preprint reports the cohort comprised patients diagnosed between 2014-2017 who had complete clinical records, pulmonary function test results, histological information, genomic data, and staging details. The dataset was divided into 70%:15%:15% training, validation, and test splits. Five model types were tuned using Hyperband, and model performance was evaluated primarily by area under the receiver operating characteristic curve (AUC), with additional metrics reported including accuracy, F1 score, precision, and recall. The authors calculated group-wise permutation importance and assessed concordance of importance rankings with the Friedman test. A Cox proportional hazards (CPH) model served as a baseline comparator.

The preprint documents that feature groups were predefined into ten categories and that Hyperband hyperparameter optimization was applied across the five candidate models. Group-wise permutation importance was used to quantify feature contributions and the Friedman test was applied to compare importance rankings across models, per the manuscript. The CPH baseline was included to provide a classical survival-analysis comparator rather than a purely discriminative benchmark.

Industry-pattern observations: using registry-scale datasets with multimodal clinical, functional, histological, and genomic inputs aligns with recent trends in prognostic modeling where richer feature sets can improve discrimination but increase the need for robust validation. Hyperparameter search frameworks such as Hyperband are commonly used in practice to make multi-model comparisons tractable on large feature sets. Permutation-based group importance combined with rank tests like Friedman provides a reproducible approach to assess feature-group consistency across model families.

Editorial analysis: this work exemplifies applied survival and prognostic modeling using national-registry data, a class of studies that can inform clinical research and risk-stratification workflows if externally validated. The presence of genomic and pulmonary function data in the KCCR-derived cohort increases potential signal for mortality prediction compared with registry datasets limited to demographics and stage alone.

For practitioners: look for the peer-reviewed publication for full performance metrics, external validation on independent cohorts, calibration analysis, and any release of model artifacts or code. Also monitor whether the authors report subgroup performance by stage, histology, or genomic subtypes and whether they provide decision-threshold guidance or explainability outputs suitable for clinical interpretation.