Learning Individual Survival Models from PanCancer Whole Transcriptome Data.

PURPOSE: Personalized medicine attempts to predict survival time for each patient, based on their individual tumor molecular profile. We investigate whether our survival learner in combination with a dimension reduction method can produce useful survival estimates for a variety of patients with cancer. EXPERIMENTAL DESIGN: This article provides a method that learns a model for predicting the survival time for individual patients with cancer from the PanCancer Atlas: given the (16,335 dimensional) gene expression profiles from 10,173 patients, each having one of 33 cancers, this method uses unsupervised nonnegative matrix factorization (NMF) to reexpress the gene expression data for each patient in terms of 100 learned NMF factors. It then feeds these 100 factors into the Multi-Task Logistic Regression (MTLR) learner to produce cancer-specific models for each of 20 cancers (with >50 uncensored instances); this produces "individual survival distributions" (ISD), which provide survival probabilities at each future time for each individual patient, which provides a patient's risk score and estimated survival time. RESULTS: Our NMF-MTLR concordance indices outperformed the VAECox benchmark by 14.9% overall. We achieved optimal survival prediction using pan-cancer NMF in combination with cancer-specific MTLR models. We provide biological interpretation of the NMF model and clinical implications of ISDs for prognosis and therapeutic response prediction. CONCLUSIONS: NMF-MTLR provides many benefits over other models: superior model discrimination, superior calibration, meaningful survival time estimates, and accurate probabilistic estimates of survival over time for each individual patient. We advocate for the adoption of these cancer survival models in clinical and research settings.

Effective hospital readmission prediction models using machine-learned features.

BACKGROUND: Hospital readmissions are one of the costliest challenges facing healthcare systems, but conventional models fail to predict readmissions well. Many existing models use exclusively manually-engineered features, which are labor intensive and dataset-specific. Our objective was to develop and evaluate models to predict hospital readmissions using derived features that are automatically generated from longitudinal data using machine learning techniques. METHODS: We studied patients discharged from acute care facilities in 2015 and 2016 in Alberta, Canada, excluding those who were hospitalized to give birth or for a psychiatric condition. We used population-level linked administrative hospital data from 2011 to 2017 to train prediction models using both manually derived features and features generated automatically from observational data. The target value of interest was 30-day all-cause hospital readmissions, with the success of prediction measured using the area under the curve (AUC) statistic. RESULTS: Data from 428,669 patients (62% female, 38% male, 27% 65 years or older) were used for training and evaluating models: 24,974 (5.83%) were readmitted within 30 days of discharge for any reason. Patients were more likely to be readmitted if they utilized hospital care more, had more physician office visits, had more prescriptions, had a chronic condition, or were 65 years old or older. The LACE readmission prediction model had an AUC of 0.66 ± 0.0064 while the machine learning model's test set AUC was 0.83 ± 0.0045, based on learning a gradient boosting machine on a combination of machine-learned and manually-derived features. CONCLUSION: Applying a machine learning model to the computer-generated and manual features improved prediction accuracy over the LACE model and a model that used only manually-derived features. Our model can be used to identify high-risk patients, for whom targeted interventions may potentially prevent readmissions.