Dynamic-DeepHit: A Deep Learning Approach for Dynamic Survival Analysis With Competing Risks Based on Longitudinal Data

The Dynamic Nature of the Socioeconomic Determinants of Cardiovascular Health: A Narrative Review

Despite decades of social epidemiologic research, health inequities remain pervasive and ubiquitous in Canada and elsewhere. One reason may be our use of socioeconomic measurement, which has often relied on single point-in-time exposures. To explore the extent to which researchers have incorporated dynamic socioeconomic measurement into cardiovascular health outcome evaluations, we performed a narrative review. We estimated the prevalence of socioeconomic longitudinal cardiovascular research studies that identified socioeconomic exposures at 2 or more points in time between the years of 2019 and 2023. We defined cardiovascular outcome studies as those that examined coronary artery disease, myocardial infarction, acute coronary syndrome, stroke, heart failure, cardiac arrhythmias, cardiac death, cardiometabolic factors, transient ischemic attacks, peripheral artery disease, or hypertension. Socioeconomic exposures included individual income, neighbourhood income, intergenerational social mobility, education, occupation, insurance status, and economic security. Seven percent of socioeconomic cardiovascular outcome studies have measured socioeconomic status at 2 or more points in time throughout the follow-up period, hypothesized mechanisms by which dynamic socioeconomic measures affected outcome focused on social mobility, accumulation, and critical period theories. Insights, implications, and future directions are discussed, in which we highlight ways in which postal code data can be better used methodologically as a dynamic socioeconomic measure. Future research must incorporate dynamic socioeconomic measurement to better inform root causes, interventions, and health-system designs if health equity is to be improved.

Enhanced osteoporotic fracture prediction in postmenopausal women using Bayesian optimization of machine learning models with genetic risk score

This study aimed to enhance the fracture risk prediction accuracy in major osteoporotic fractures (MOFs) and hip fractures (HFs) by integrating genetic profiles, machine learning (ML) techniques, and Bayesian optimization. The genetic risk score (GRS), derived from 1,103 risk single nucleotide polymorphisms (SNPs) from genome-wide association studies (GWAS), was formulated for 25,772 postmenopausal women from the Women's Health Initiative dataset. We developed four ML models: Support Vector Machine (SVM), Random Forest, XGBoost, and Artificial Neural Network (ANN) for binary fracture outcome and 10-year fracture risk prediction. GRS and FRAX clinical risk factors (CRFs) were used as predictors. Death as a competing risk was accounted for in ML models for time-to-fracture data. ML models were subsequently fine-tuned through Bayesian optimization, which displayed marked superiority over traditional grid search. Evaluation of the models' performance considered an array of metrics such as accuracy, weighted F1 Score, the area under the precision-recall curve (PRAUC), and the area under the receiver operating characteristic curve (AUC) for binary fracture predictions, and the C-index, Brier score, and dynamic mean AUC over a 10-year follow-up period for fracture risk predictions. We found that GRS-integrated XGBoost with Bayesian optimization is the most effective model, with an accuracy of 91.2% (95% CI: 90.4-92.0%) and an AUC of 0.739 (95% CI: 0.731-0.746) in MOF binary predictions. For 10-year fracture risk modeling, the XGBoost model attained a C-index of 0.795 (95% CI: 0.783-0.806) and a mean dynamic AUC of 0.799 (95% CI: 0.788-0.809). Compared to FRAX, the XGBoost model exhibited a categorical net reclassification improvement (NRI) of 22.6% (P = .004). A sensitivity analysis, which included BMD but lacked GRS, reaffirmed these findings. Furthermore, portability tests in diverse non-European groups, including Asians and African Americans, underscored the model's robustness and adaptability. This study accentuates the potential of combining genetic insights and optimized ML in strengthening fracture predictions, heralding new preventive strategies for postmenopausal women.

Exploring factors influencing time from dispatch to unit availability according to the transport decision in the pre-hospital setting: an exploratory study

BACKGROUND

Efficient resource distribution is important. Despite extensive research on response timings within ambulance services, nuances of time from unit dispatch to becoming available still need to be explored. This study aimed to identify the determinants of the duration between ambulance dispatch and readiness to respond to the next case according to the patients' transport decisions.

METHODS

Time from ambulance dispatch to availability (TDA) analysis according to the patients' transport decision (Transport versus Non-Transport) was conducted using R-Studio™ for a data set of 93,712 emergency calls managed by a Middle Eastern ambulance service from January to May 2023. Log-transformed Hazard Ratios (HR) were examined across diverse parameters. A Cox regression model was utilised to determine the influence of variables on TDA. Kaplan-Meier curves discerned potential variances in the time elapsed for both cohorts based on demographics and clinical indicators. A competing risk analysis assessed the probabilities of distinct outcomes occurring.

RESULTS

The median duration of elapsed TDA was 173 min for the transported patients and 73 min for those not transported. The HR unveiled Significant associations in various demographic variables. The Kaplan-Meier curves revealed variances in TDA across different nationalities and age categories. In the competing risk analysis, the 'Not Transported' group demonstrated a higher incidence of prolonged TDA than the 'Transported' group at specified time points.

CONCLUSIONS

Exploring TDA offers a novel perspective on ambulance services' efficiency. Though promising, the findings necessitate further exploration across diverse settings, ensuring broader applicability. Future research should consider a comprehensive range of variables to fully harness the utility of this period as a metric for healthcare excellence.

Survival Analysis of High-Dimensional Data With Graph Convolutional Networks and Geometric Graphs

This article proposes a survival model based on graph convolutional networks (GCNs) with geometric graphs directly constructed from high-dimensional features. First, we clarify that the graphs used in GCNs play an important role in processing the relational information of samples, and the graphs that align well with the underlying data structure could be beneficial for survival analysis. Second, we show that sparse geometric graphs derived from high-dimensional data are more favorable compared with dense graphs when used in GCNs for survival analysis. Third, from this insight, we propose a model for survival analysis based on GCNs. By using multiple sparse geometric graphs and a proposed sequential forward floating selection algorithm, the new model is able to simultaneously perform survival analysis and unveil the local neighborhoods of samples. The experimental results on real-world datasets show that the proposed survival analysis approach based on GCNs outperforms a variety of existing methods and indicate that geometric graphs can aid survival analysis of high-dimensional data.

Developing deep learning-based strategies to predict the risk of hepatocellular carcinoma among patients with nonalcoholic fatty liver disease from electronic health records

OBJECTIVE

The accuracy of deep learning models for many disease prediction problems is affected by time-varying covariates, rare incidence, covariate imbalance and delayed diagnosis when using structured electronic health records data. The situation is further exasperated when predicting the risk of one disease on condition of another disease, such as the hepatocellular carcinoma risk among patients with nonalcoholic fatty liver disease due to slow, chronic progression, the scarce of data with both disease conditions and the sex bias of the diseases. The goal of this study is to investigate the extent to which the aforementioned issues influence deep learning performance, and then devised strategies to tackle these challenges. These strategies were applied to improve hepatocellular carcinoma risk prediction among patients with nonalcoholic fatty liver disease.

METHODS

We evaluated two representative deep learning models in the task of predicting the occurrence of hepatocellular carcinoma in a cohort of patients with nonalcoholic fatty liver disease (n = 220,838) from a national EHR database. The disease prediction task was carefully formulated as a classification problem while taking censorship and the length of follow-up into consideration.

RESULTS

We developed a novel backward masking scheme to deal with the issue of delayed diagnosis which is very common in EHR data analysis and evaluate how the length of longitudinal information after the index date affects disease prediction. We observed that modeling time-varying covariates improved the performance of the algorithms and transfer learning mitigated reduced performance caused by the lack of data. In addition, covariate imbalance, such as sex bias in data impaired performance. Deep learning models trained on one sex and evaluated in the other sex showed reduced performance, indicating the importance of assessing covariate imbalance while preparing data for model training.

CONCLUSIONS

The strategies developed in this work can significantly improve the performance of hepatocellular carcinoma risk prediction among patients with nonalcoholic fatty liver disease. Furthermore, our novel strategies can be generalized to apply to other disease risk predictions using structured electronic health records, especially for disease risks on condition of another disease.

Optimizing dynamic predictions from joint models using super learning

Joint models for longitudinal and time-to-event data are often employed to calculate dynamic individualized predictions used in numerous applications of precision medicine. Two components of joint models that influence the accuracy of these predictions are the shape of the longitudinal trajectories and the functional form linking the longitudinal outcome history to the hazard of the event. Finding a single well-specified model that produces accurate predictions for all subjects and follow-up times can be challenging, especially when considering multiple longitudinal outcomes. In this work, we use the concept of super learning and avoid selecting a single model. In particular, we specify a weighted combination of the dynamic predictions calculated from a library of joint models with different specifications. The weights are selected to optimize a predictive accuracy metric using V-fold cross-validation. We use as predictive accuracy measures the expected quadratic prediction error and the expected predictive cross-entropy. In a simulation study, we found that the super learning approach produces results very similar to the Oracle model, which was the model with the best performance in the test datasets. All proposed methodology is implemented in the freely available R package JMbayes2.

Hybridizing mechanistic mathematical modeling with deep learning methods to predict individual cancer patient survival after immune checkpoint inhibitor therapy

We present a study where predictive mechanistic modeling is used in combination with deep learning methods to predict individual patient survival probabilities under immune checkpoint inhibitor (ICI) therapy. This hybrid approach enables prediction based on both measures that are calculable from mechanistic models (but may not be directly measurable in the clinic) and easily measurable quantities or characteristics (that are not always readily incorporated into predictive mechanistic models). The mechanistic model we have applied here can predict tumor response from CT or MRI imaging based on key mechanisms underlying checkpoint inhibitor therapy, and in the present work, its parameters were combined with readily-available clinical measures from 93 patients into a hybrid training set for a deep learning time-to-event predictive model. Analysis revealed that training an artificial neural network with both mechanistic modeling-derived and clinical measures achieved higher per-patient predictive accuracy based on event-time concordance, Brier score, and negative binomial log-likelihood-based criteria than when only mechanistic model-derived values or only clinical data were used. Feature importance analysis revealed that both clinical and model-derived parameters play prominent roles in neural network decision making, and in increasing prediction accuracy, further supporting the advantage of our hybrid approach. We anticipate that many existing mechanistic models may be hybridized with deep learning methods in a similar manner to improve predictive accuracy through addition of additional data that may not be readily implemented in mechanistic descriptions.

Added prognostic value of 3D deep learning-derived features from preoperative MRI for adult-type diffuse gliomas

BACKGROUND

To investigate the prognostic value of spatial features from whole-brain MRI using a three-dimensional (3D) convolutional neural network for adult-type diffuse gliomas.

METHODS

In a retrospective, multicenter study, 1925 diffuse glioma patients were enrolled from 5 datasets: SNUH (n = 708), UPenn (n = 425), UCSF (n = 500), TCGA (n = 160), and Severance (n = 132). The SNUH and Severance datasets served as external test sets. Precontrast and postcontrast 3D T1-weighted, T2-weighted, and T2-FLAIR images were processed as multichannel 3D images. A 3D-adapted SE-ResNeXt model was trained to predict overall survival. The prognostic value of the deep learning-based prognostic index (DPI), a spatial feature-derived quantitative score, and established prognostic markers were evaluated using Cox regression. Model evaluation was performed using the concordance index (C-index) and Brier score.

RESULTS

The MRI-only median DPI survival prediction model achieved C-indices of 0.709 and 0.677 (BS = 0.142 and 0.215) and survival differences (P < 0.001 and P = 0.002; log-rank test) for the SNUH and Severance datasets, respectively. Multivariate Cox analysis revealed DPI as a significant prognostic factor, independent of clinical and molecular genetic variables: hazard ratio = 0.032 and 0.036 (P < 0.001 and P = 0.004) for the SNUH and Severance datasets, respectively. Multimodal prediction models achieved higher C-indices than models using only clinical and molecular genetic variables: 0.783 vs. 0.774, P = 0.001, SNUH; 0.766 vs. 0.748, P = 0.023, Severance.

CONCLUSIONS

The global morphologic feature derived from 3D CNN models using whole-brain MRI has independent prognostic value for diffuse gliomas. Combining clinical, molecular genetic, and imaging data yields the best performance.

Development and Application of a Novel Machine Learning Model Predicting Pancreatic Cancer-Specific Mortality

Precise prognostication is vital for guiding treatment decisions in people diagnosed with pancreatic cancer. Existing models depend on predetermined variables, constraining their effectiveness. Our objective was to explore a novel machine learning approach to enhance a prognostic model for predicting pancreatic cancer-specific mortality and, subsequently, to assess its performance against Cox regression models. Datasets were retrospectively collected and analyzed for 9,752 patients diagnosed with pancreatic cancer and with surgery performed. The primary outcomes were the mortality of patients with pancreatic carcinoma at one year, three years, and five years. Model discrimination was assessed using the concordance index (C-index), and calibration was assessed using Brier scores. The Survival Quilts model was compared with Cox regression models in clinical use, and decision curve analysis was done. The Survival Quilts model demonstrated robust discrimination for one-year (C-index 0.729), three-year (C-index 0.693), and five-year (C-index 0.672) pancreatic cancer-specific mortality. In comparison to Cox models, the Survival Quilts models exhibited a higher C-index up to 32 months but displayed inferior performance after 33 months. A subgroup analysis was conducted, revealing that within the subset of individuals without metastasis, the Survival Quilts models showcased a significant advantage over the Cox models. In the cohort with metastatic pancreatic cancer, Survival Quilts outperformed the Cox model before 24 months but exhibited a weaker performance after 25 months. This study has developed and validated a novel machine learning-based Survival Quilts model to predict pancreatic cancer-specific mortality that outperforms the Cox regression model.

Utility of multimodal longitudinal imaging data for dynamic prediction of cardiovascular and renal disease: the CARDIA study

Background

Medical examinations contain repeatedly measured data from multiple visits, including imaging variables collected from different modalities. However, the utility of such data for the prediction of time-to-event is unknown, and only a fraction of the data is typically used for risk prediction. We hypothesized that multimodal longitudinal imaging data could improve dynamic disease prognosis of cardiovascular and renal disease (CVRD).

Methods

In a multi-centered cohort of 5,114 CARDIA participants, we included 166 longitudinal imaging variables from five imaging modalities: Echocardiography (Echo), Cardiac and Abdominal Computed Tomography (CT), Dual-Energy x-ray Absorptiometry (DEXA), Brain Magnetic Resonance Imaging (MRI) collected from young adulthood to mid-life over 30 years (1985-2016) to perform dynamic survival analysis of CVRD events using machine learning dynamic survival analysis (Dynamic-DeepHit, LTRCforest, and Extended Cox for Time-varying Covariates). Risk probabilities were continuously updated as new data were collected. Model performance was assessed using integrated AUC and C-index and compared to traditional risk factors.

Results

Longitudinal imaging data, even when being irregularly collected with high missing rates, improved CVRD dynamic prediction (0.03 in integrated AUC, up to 0.05 in C-index compared to traditional risk factors; best model's C-index = 0.80-0.83 up to 20 years from baseline) from young adulthood followed up to midlife. Among imaging variables, Echo and CT variables contributed significantly to improved risk estimation. Echo measured in early adulthood predicted midlife CVRD risks almost as well as Echo measured 10-15 years later (0.01 C-index difference). The most recent CT exam provided the most accurate prediction for short-term risk estimation. Brain MRI markers provided additional information from cardiac Echo and CT variables that led to a slightly improved prediction.

Conclusions

Longitudinal multimodal imaging data readily collected from follow-up exams can improve CVRD dynamic prediction. Echocardiography measured early can provide a good long-term risk estimation, while CT/calcium scoring variables carry atherosclerotic signatures that benefit more immediate risk assessment starting in middle-age.

Incorporating longitudinal history of risk factors into atherosclerotic cardiovascular disease risk prediction using deep learning

It is increasingly clear that longitudinal risk factor levels and trajectories are related to risk for atherosclerotic cardiovascular disease (ASCVD) above and beyond single measures. Currently used in clinical care, the Pooled Cohort Equations (PCE) are based on regression methods that predict ASCVD risk based on cross-sectional risk factor levels. Deep learning (DL) models have been developed to incorporate longitudinal data for risk prediction but its benefit for ASCVD risk prediction relative to the traditional Pooled Cohort Equations (PCE) remain unknown. Our study included 15,565 participants from four cardiovascular disease cohorts free of baseline ASCVD who were followed for adjudicated ASCVD. Ten-year ASCVD risk was calculated in the training set using our benchmark, the PCE, and a longitudinal DL model, Dynamic-DeepHit. Predictors included those incorporated in the PCE: sex, race, age, total cholesterol, high density lipid cholesterol, systolic and diastolic blood pressure, diabetes, hypertension treatment and smoking. The discrimination and calibration performance of the two models were evaluated in an overall hold-out testing dataset. Of the 15,565 participants in our dataset, 2170 (13.9%) developed ASCVD. The performance of the longitudinal DL model that incorporated 8 years of longitudinal risk factor data improved upon that of the PCE [AUROC: 0.815 (CI 0.782-0.844) vs 0.792 (CI 0.760-0.825)] and the net reclassification index was 0.385. The brier score for the DL model was 0.0514 compared with 0.0542 in the PCE. Incorporating longitudinal risk factors in ASCVD risk prediction using DL can improve model discrimination and calibration.

Deep learning-assisted survival prognosis in renal cancer: A CT scan-based personalized approach

This paper presents a deep learning (DL) approach for predicting survival probabilities of renal cancer patients based solely on preoperative CT imaging. The proposed approach consists of two networks: a classifier- and a survival- network. The classifier attempts to extract features from 3D CT scans to predict the ISUP grade of Renal cell carcinoma (RCC) tumors, as defined by the International Society of Urological Pathology (ISUP). Our classifier is a 3D convolutional neural network to avoid losing crucial information on the interconnection of slides in 3D images. We employ multiple procedures, including image augmentation, preprocessing, and concatenation, to improve the performance of the classifier. Given the strong correlation between ISUP grading and renal cancer prognosis in the clinical context, we use the ISUP grading features extracted by the classifier as the input to the survival network. By leveraging this clinical association and the classifier network, we are able to model our survival analysis using a simple DL-based network. We adopt a discrete LogisticHazard-based loss to extract intrinsic survival characteristics of RCC tumors from CT images. This allows us to build a completely parametric survival model that varies with patients' tumor characteristics and predicts non-proportional survival probability curves for different patients. Our results demonstrated that the proposed method could predict the future course of renal cancer with reasonable accuracy from the CT scans. The proposed method obtained an average concordance index of 0.72, an integrated Brier score of 0.15, and an area under the curve value of 0.71 on the test cohorts.

Dynamic prediction of malignant ventricular arrhythmias using neural networks in patients with an implantable cardioverter-defibrillator

BACKGROUND

Risk stratification for ventricular arrhythmias currently relies on static measurements that fail to adequately capture dynamic interactions between arrhythmic substrate and triggers over time. We trained and internally validated a dynamic machine learning (ML) model and neural network that extracted features from longitudinally collected electrocardiograms (ECG), and used these to predict the risk of malignant ventricular arrhythmias.

METHODS

A multicentre study in patients implanted with an implantable cardioverter-defibrillator (ICD) between 2007 and 2021 in two academic hospitals was performed. Variational autoencoders (VAEs), which combine neural networks with variational inference principles, and can learn patterns and structure in data without explicit labelling, were trained to encode the mean ECG waveforms from the limb leads into 16 variables. Supervised dynamic ML models using these latent ECG representations and clinical baseline information were trained to predict malignant ventricular arrhythmias treated by the ICD. Model performance was evaluated on a hold-out set, using time-dependent receiver operating characteristic (ROC) and calibration curves.

FINDINGS

2942 patients (61.7 ± 13.9 years, 25.5% female) were included, with a total of 32,129 ECG recordings during a mean follow-up of 43.9 ± 35.9 months. The mean time-varying area under the ROC curve for the dynamic model was 0.738 ± 0.07, compared to 0.639 ± 0.03 for a static (i.e. baseline-only model). Feature analyses indicated dynamic changes in latent ECG representations, particularly those affecting the T-wave morphology, were of highest importance for model predictions.

INTERPRETATION

Dynamic ML models and neural networks effectively leverage routinely collected longitudinal ECG recordings for personalised and updated predictions of malignant ventricular arrhythmias, outperforming static models.

FUNDING

This publication is part of the project DEEP RISK ICD (with project number 452019308) of the research programme Rubicon which is (partly) financed by the Dutch Research Council (NWO). This research is partly funded by the Amsterdam Cardiovascular Sciences (personal grant F.V.Y.T).

Sample size and predictive performance of machine learning methods with survival data: A simulation study

Prediction models are increasingly developed and used in diagnostic and prognostic studies, where the use of machine learning (ML) methods is becoming more and more popular over traditional regression techniques. For survival outcomes the Cox proportional hazards model is generally used and it has been proven to achieve good prediction performances with few strong covariates. The possibility to improve the model performance by including nonlinearities, covariate interactions and time-varying effects while controlling for overfitting must be carefully considered during the model building phase. On the other hand, ML techniques are able to learn complexities from data at the cost of hyper-parameter tuning and interpretability. One aspect of special interest is the sample size needed for developing a survival prediction model. While there is guidance when using traditional statistical models, the same does not apply when using ML techniques. This work develops a time-to-event simulation framework to evaluate performances of Cox regression compared, among others, to tuned random survival forest, gradient boosting, and neural networks at varying sample sizes. Simulations were based on replications of subjects from publicly available databases, where event times were simulated according to a Cox model with nonlinearities on continuous variables and time-varying effects and on the SEER registry data.

Private Continuous Survival Analysis with Distributed Multi-Site Data

Effective disease surveillance systems require large-scale epidemiological data to improve health outcomes and quality of care for the general population. As data may be limited within a single site, multi-site data (e.g., from a number of local/regional health systems) need to be considered. Leveraging distributed data across multiple sites for epidemiological analysis poses significant challenges. Due to the sensitive nature of epidemiological data, it is imperative to design distributed solutions that provide strong privacy protections. Current privacy solutions often assume a central site, which is responsible for aggregating the distributed data and applying privacy protection before sharing the results (e.g., aggregation via secure primitives and differential privacy for sharing aggregate results). However, identifying such a central site may be difficult in practice and relying on a central site may introduce potential vulnerabilities (e.g., single point of failure). Furthermore, to support clinical interventions and inform policy decisions in a timely manner, epidemiological analysis need to reflect dynamic changes in the data. Yet, existing distributed privacy-protecting approaches were largely designed for static data (e.g., one-time data sharing) and cannot fulfill dynamic data requirements. In this work, we propose a privacy-protecting approach that supports the sharing of dynamic epidemiological analysis and provides strong privacy protection in a decentralized manner. We apply our solution in continuous survival analysis using the Kaplan-Meier estimation model while providing differential privacy protection. Our evaluations on a real dataset containing COVID-19 cases show that our method provides highly usable results.

Comparison of MRI radiomics-based machine learning survival models in predicting prognosis of glioblastoma multiforme

Objective

To compare the performance of radiomics-based machine learning survival models in predicting the prognosis of glioblastoma multiforme (GBM) patients.

Methods

131 GBM patients were included in our study. The traditional Cox proportional-hazards (CoxPH) model and four machine learning models (SurvivalTree, Random survival forest (RSF), DeepSurv, DeepHit) were constructed, and the performance of the five models was evaluated using the C-index.

Results

After the screening, 1792 radiomics features were obtained. Seven radiomics features with the strongest relationship with prognosis were obtained following the application of the least absolute shrinkage and selection operator (LASSO) regression. The CoxPH model demonstrated that age (HR = 1.576, p = 0.037), Karnofsky performance status (KPS) score (HR = 1.890, p = 0.006), radiomics risk score (HR = 3.497, p = 0.001), and radiomics risk level (HR = 1.572, p = 0.043) were associated with poorer prognosis. The DeepSurv model performed the best among the five models, obtaining C-index of 0.882 and 0.732 for the training and test set, respectively. The performances of the other four models were lower: CoxPH (0.663 training set / 0.635 test set), SurvivalTree (0.702/0.655), RSF (0.735/0.667), DeepHit (0.608/0.560).

Conclusion

This study confirmed the superior performance of deep learning algorithms based on radiomics relative to the traditional method in predicting the overall survival of GBM patients; specifically, the DeepSurv model showed the best predictive ability.

Prognostic significance of cyclin D1 expression pattern in HPV-negative oral and oropharyngeal carcinoma: A deep-learning approach

BACKGROUND

We aimed to establish image recognition and survival prediction models using a novel scoring system of cyclin D1 expression pattern in patients with human papillomavirus-negative oral or oropharyngeal squamous cell carcinoma.

METHODS

The clinicopathological data of 610 patients with human papillomavirus-negative oral/oropharyngeal squamous cell carcinoma were analyzed retrospectively. Cox univariate and multivariate risk regression analyses were performed to compare cyclin D1 expression pattern scoring with the traditional scoring method-cyclin D1 expression level scoring-in relation to patients' overall and progression-free survival. An image recognition model employing the cyclin D1 expression pattern scoring system was established by YOLOv5 algorithms. From this model, two independent survival prediction models were established using the DeepHit and DeepSurv algorithms.

RESULTS

Cyclin D1 had three expression patterns in oral and oropharyngeal squamous cell carcinoma cancer nests. Superior to cyclin D1 expression level scoring, cyclin D1 expression pattern scoring was significantly correlated with the prognosis of patients with oral squamous cell carcinoma (p < 0.0001) and oropharyngeal squamous cell carcinoma (p < 0.05). Moreover, it was an independent prognostic risk factor in both oral squamous cell carcinoma (p < 0.0001) and oropharyngeal squamous cell carcinoma (p < 0.05). The cyclin D1 expression pattern-derived image recognition model showed an average test set accuracy of 78.48% ± 4.31%. In the overall survival prediction models, the average concordance indices of the test sets established by DeepSurv and DeepHit were 0.71 ± 0.02 and 0.70 ± 0.01, respectively.

CONCLUSION

Combined with the image recognition model of the cyclin D1 expression pattern, the survival prediction model had a relatively good prediction effect on the overall survival prognosis of patients with human papillomavirus-negative oral or oropharyngeal squamous cell carcinoma.

Bridging the Worlds of Pharmacometrics and Machine Learning

Precision medicine requires individualized modeling of disease and drug dynamics, with machine learning-based computational techniques gaining increasing popularity. The complexity of either field, however, makes current pharmacological problems opaque to machine learning practitioners, and state-of-the-art machine learning methods inaccessible to pharmacometricians. To help bridge the two worlds, we provide an introduction to current problems and techniques in pharmacometrics that ranges from pharmacokinetic and pharmacodynamic modeling to pharmacometric simulations, model-informed precision dosing, and systems pharmacology, and review some of the machine learning approaches to address them. We hope this would facilitate collaboration between experts, with complementary strengths of principled pharmacometric modeling and flexibility of machine learning leading to synergistic effects in pharmacological applications.

Incorporating longitudinal history of risk factors into atherosclerotic cardiovascular disease risk prediction using deep learning

Background

It is increasingly clear that longitudinal risk factor levels and trajectories are related to risk for atherosclerotic cardiovascular disease (ASCVD) above and beyond single measures. Currently used in clinical care, the Pooled Cohort Equations (PCE) are based on regression methods that predict ASCVD risk based on cross-sectional risk factor levels. Deep learning (DL) models have been developed to incorporate longitudinal data for risk prediction but its benefit for ASCVD risk prediction relative to the traditional Pooled Cohort Equations (PCE) remain unknown.

Objective

To develop a ASCVD risk prediction model that incorporates longitudinal risk factors using deep learning.

Methods

Our study included 15,565 participants from four cardiovascular disease cohorts free of baseline ASCVD who were followed for adjudicated ASCVD. Ten-year ASCVD risk was calculated in the training set using our benchmark, the PCE, and a longitudinal DL model, Dynamic-DeepHit. Predictors included those incorporated in the PCE: sex, race, age, total cholesterol, high density lipid cholesterol, systolic and diastolic blood pressure, diabetes, hypertension treatment and smoking. The discrimination and calibration performance of the two models were evaluated in an overall hold-out testing dataset.

Results

Of the 15,565 participants in our dataset, 2,170 (13.9%) developed ASCVD. The performance of the longitudinal DL model that incorporated 8 years of longitudinal risk factor data improved upon that of the PCE [AUROC: 0.815 (CI: 0.782-0.844) vs 0.792 (CI: 0.760-0.825)] and the net reclassification index was 0.385. The brier score for the DL model was 0.0514 compared with 0.0542 in the PCE.

Conclusion

Incorporating longitudinal risk factors in ASCVD risk prediction using DL can improve model discrimination and calibration.

A Dynamic Prediction Model Supporting Individual Life Expectancy Prediction Based on Longitudinal Time-Dependent Covariates

In the field of clinical chronic diseases, common prediction results (such as survival rate) and effect size hazard ratio (HR) are relative indicators, resulting in more abstract information. However, clinicians and patients are more interested in simple and intuitive concepts of (survival) time, such as how long a patient may live or how much longer a patient in a treatment group will live. In addition, due to the long follow-up time, resulting in generation of longitudinal time-dependent covariate information, patients are interested in how long they will survive at each follow-up visit. In this study, based on a time scale indicator-restricted mean survival time (RMST)-we proposed a dynamic RMST prediction model by considering longitudinal time-dependent covariates and utilizing joint model techniques. The model can describe the change trajectory of longitudinal time-dependent covariates and predict the average survival times of patients at different time points (such as follow-up visits). Simulation studies through Monte Carlo cross-validation showed that the dynamic RMST prediction model was superior to the static RMST model. In addition, the dynamic RMST prediction model was applied to a primary biliary cirrhosis (PBC) population to dynamically predict the average survival times of the patients, and the average C-index of the internal validation of the model reached 0.81, which was better than that of the static RMST regression. Therefore, the proposed dynamic RMST prediction model has better performance in prediction and can provide a scientific basis for clinicians and patients to make clinical decisions.

Germline mechanisms of immunotherapy toxicities in the era of genome-wide association studies

Cancer immunotherapy has revolutionized the treatment of advanced cancers and is quickly becoming an option for early-stage disease. By reactivating the host immune system, immunotherapy harnesses patients' innate defenses to eradicate the tumor. By putatively similar mechanisms, immunotherapy can also substantially increase the risk of toxicities or immune-related adverse events (irAEs). Severe irAEs can lead to hospitalization, treatment discontinuation, lifelong immune complications, or even death. Many irAEs present with similar symptoms to heritable autoimmune diseases, suggesting that germline genetics may contribute to their onset. Recently, genome-wide association studies (GWAS) of irAEs have identified common germline associations and putative mechanisms, lending support to this hypothesis. A wide range of well-established GWAS methods can potentially be harnessed to understand the etiology of irAEs specifically and immunotherapy outcomes broadly. This review summarizes current findings regarding germline effects on immunotherapy outcomes and discusses opportunities and challenges for leveraging germline genetics to understand, predict, and treat irAEs.

Multi-task banded regression model: A novel individual survival analysis model for breast cancer

PURPOSE

To reveal the hazard probability of individual breast cancer patients, a multi-task banded regression model is proposed for individual survival analysis of breast cancer.

METHODS

A banded verification matrix is designed to construct the response transform function of the proposed multi-task banded regression model, which can solve the repeated switching of survival rate. A martingale process is introduced to construct different nonlinear regressions for different survival subintervals. The concordance index (C-index) is used to compare the proposed model with Cox proportional hazards (CoxPH) models and previous multi-task regression models.

RESULTS

Two commonly-used breast cancer datasets are employed to validate the proposed model. Specifically, the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) includes 1981 breast cancer patients, of which 57.7% died of breast cancer. The Rotterdam & German Breast Cancer Study Group (GBSG) includes 1546 patients with lymph node-positive breast cancer in a randomized clinical trial, of which 44.4% died. Experimental results indicate that the proposed model is superior to some existing models for overall and individual survival analysis of breast cancer, with the C-index of 0.6786 for the GBSG and 0.6701 for the METABRIC.

CONCLUSION

The superiority of the proposed model can be contributed to three novel ideas. One is that a banded verification matrix can band the response of the survival process. Second, the martingale process can construct different nonlinear regressions for different survival subintervals. Third, the novel loss can adapt the model to making the multi-task regression similar to the real survival process.

Predictive Models for Health Deterioration: Understanding Disease Pathways for Personalized Medicine

Artificial intelligence (AI) and machine learning (ML) methods are currently widely employed in medicine and healthcare. A PubMed search returns more than 100,000 articles on these topics published between 2018 and 2022 alone. Notwithstanding several recent reviews in various subfields of AI and ML in medicine, we have yet to see a comprehensive review around the methods' use in longitudinal analysis and prediction of an individual patient's health status within a personalized disease pathway. This review seeks to fill that gap. After an overview of the AI and ML methods employed in this field and of specific medical applications of models of this type, the review discusses the strengths and limitations of current studies and looks ahead to future strands of research in this field. We aim to enable interested readers to gain a detailed impression of the research currently available and accordingly plan future work around predictive models for deterioration in health status.

High-Dimensional Survival Analysis: Methods and Applications

In the era of precision medicine, time-to-event outcomes such as time to death or progression are routinely collected, along with high-throughput covariates. These high-dimensional data defy classical survival regression models, which are either infeasible to fit or likely to incur low predictability due to over-fitting. To overcome this, recent emphasis has been placed on developing novel approaches for feature selection and survival prognostication. We will review various cutting-edge methods that handle survival outcome data with high-dimensional predictors, highlighting recent innovations in machine learning approaches for survival prediction. We will cover the statistical intuitions and principles behind these methods and conclude with extensions to more complex settings, where competing events are observed. We exemplify these methods with applications to the Boston Lung Cancer Survival Cohort study, one of the largest cancer epidemiology cohorts investigating the complex mechanisms of lung cancer.

Deep-learning-based prognostic modeling for incident heart failure in patients with diabetes using electronic health records: A retrospective cohort study

Patients with type 2 diabetes mellitus (T2DM) have more than twice the risk of developing heart failure (HF) compared to patients without diabetes. The present study is aimed to build an artificial intelligence (AI) prognostic model that takes in account a large and heterogeneous set of clinical factors and investigates the risk of developing HF in diabetic patients. We carried out an electronic health records- (EHR-) based retrospective cohort study that included patients with cardiological clinical evaluation and no previous diagnosis of HF. Information consists of features extracted from clinical and administrative data obtained as part of routine medical care. The primary endpoint was diagnosis of HF (during out-of-hospital clinical examination or hospitalization). We developed two prognostic models using (1) elastic net regularization for Cox proportional hazard model (COX) and (2) a deep neural network survival method (PHNN), in which a neural network was used to represent a non-linear hazard function and explainability strategies are applied to estimate the influence of predictors on the risk function. Over a median follow-up of 65 months, 17.3% of the 10,614 patients developed HF. The PHNN model outperformed COX both in terms of discrimination (c-index 0.768 vs 0.734) and calibration (2-year integrated calibration index 0.008 vs 0.018). The AI approach led to the identification of 20 predictors of different domains (age, body mass index, echocardiographic and electrocardiographic features, laboratory measurements, comorbidities, therapies) whose relationship with the predicted risk correspond to known trends in the clinical practice. Our results suggest that prognostic models for HF in diabetic patients may improve using EHRs in combination with AI techniques for survival analysis, which provide high flexibility and better performance with respect to standard approaches.

Modeling comorbidity of chronic diseases using coupled hidden Markov model with bivariate discrete copula

A range of chronic diseases have a significant influence on each other and share common risk factors. Comorbidity, which shows the existence of two or more diseases interacting or triggering each other, is an important measure for actuarial valuations. The main proposal of the study is to model parallel interacting processes describing two or more chronic diseases by a combination of hidden Markov theory and copula function. This study introduces a coupled hidden Markov model with the bivariate discrete copula function in the hidden process. To estimate the parameters of the model and deal with the numerical intractability of the log-likelihood, we use a variational expectation maximization algorithm. To perform the variational expectation maximization algorithm, a lower bound of the model's log-likelihood is defined, and estimators of the parameters are computed in the M-part. A possible numerical underflow occurring in the computation of forward-backward probabilities is solved. The simulation study is conducted for two different levels of association to assess the performance of the proposed model, resulting in satisfactory findings. The proposed model was applied to hospital appointment data from a private hospital. The model defines the dependency structure of unobserved disease data and its dynamics. The application results demonstrate that the model is useful for investigating disease comorbidity when only population dynamics over time and no clinical data are available.

Multivariate longitudinal data for survival analysis of cardiovascular event prediction in young adults: insights from a comparative explainable study

BACKGROUND

Multivariate longitudinal data are under-utilized for survival analysis compared to cross-sectional data (CS - data collected once across cohort). Particularly in cardiovascular risk prediction, despite available methods of longitudinal data analysis, the value of longitudinal information has not been established in terms of improved predictive accuracy and clinical applicability.

METHODS

We investigated the value of longitudinal data over and above the use of cross-sectional data via 6 distinct modeling strategies from statistics, machine learning, and deep learning that incorporate repeated measures for survival analysis of the time-to-cardiovascular event in the Coronary Artery Risk Development in Young Adults (CARDIA) cohort. We then examined and compared the use of model-specific interpretability methods (Random Survival Forest Variable Importance) and model-agnostic methods (SHapley Additive exPlanation (SHAP) and Temporal Importance Model Explanation (TIME)) in cardiovascular risk prediction using the top-performing models.

RESULTS

In a cohort of 3539 participants, longitudinal information from 35 variables that were repeatedly collected in 6 exam visits over 15 years improved subsequent long-term (17 years after) risk prediction by up to 8.3% in C-index compared to using baseline data (0.78 vs. 0.72), and up to approximately 4% compared to using the last observed CS data (0.75). Time-varying AUC was also higher in models using longitudinal data (0.86-0.87 at 5 years, 0.79-0.81 at 10 years) than using baseline or last observed CS data (0.80-0.86 at 5 years, 0.73-0.77 at 10 years). Comparative model interpretability analysis revealed the impact of longitudinal variables on model prediction on both the individual and global scales among different modeling strategies, as well as identifying the best time windows and best timing within that window for event prediction. The best strategy to incorporate longitudinal data for accuracy was time series massive feature extraction, and the easiest interpretable strategy was trajectory clustering.

CONCLUSION

Our analysis demonstrates the added value of longitudinal data in predictive accuracy and epidemiological utility in cardiovascular risk survival analysis in young adults via a unified, scalable framework that compares model performance and explainability. The framework can be extended to a larger number of variables and other longitudinal modeling methods.

TRIAL REGISTRATION

ClinicalTrials.gov Identifier: NCT00005130, Registration Date: 26/05/2000.

Spatio-temporally smoothed deep survival neural network

The analysis of registry data has important implications for cancer monitoring, control, and treatment. In such analysis, (semi)parametric models, such as the Cox Proportional Hazards model, have been routinely adopted. In recent years, deep neural network (DNN) has been shown to excel in many fields with its flexibility and superior prediction performance, and it has been applied to the analysis of cancer survival data. Cancer registry data usually has a broad spatial and temporal coverage, leading to significant heterogeneity. Published studies have suggested that it is not sensible to fit one model for all spatial and temporal locations combined. On the other hand, it is inefficient to fit one model for each spatial/temporal location separately. Motivated by such considerations, in this study, we develop a spatio-temporally smoothed DNN approach for the analysis of cancer registry data with a (censored) survival outcome. This approach can accommodate the significant differences across time and space, while recognizing that the spatial and temporal changes are smooth. It is effectively realized via cutting-edge optimization techniques. To draw more definitive conclusions, we also develop an approach for assessing the importance of each individual input variable. Data on head and neck cancer (HNC) and pancreatic cancer from the Surveillance, Epidemiology, and End Results (SEER) database is analyzed. Compared to direct competitors, the proposed approach leads to network architectures that are smoother. Evaluated using the time-dependent Concordance-Index, it has a better prediction performance. The important variables are also biomedically sensible. Overall, this study can deliver a new and effective tool for deciphering cancer survival at the population level.

TERTIAN: Clinical Endpoint Prediction in ICU via Time-Aware Transformer-Based Hierarchical Attention Network

Accurately predicting the clinical endpoint in ICU based on the patient's electronic medical records (EMRs) is essential for the timely treatment of critically ill patients and allocation of medical resources. However, the patient's EMRs usually consist of a large amount of heterogeneous multivariate time series data such as laboratory tests and vital signs, which are produced irregularly. Most existing methods fail to effectively model the time irregularity inherent in longitudinal patient medical records and capture the interrelationships among different types of data. To tackle these limitations, we propose a novel time-aware transformer-based hierarchical attention network (TERTIAN) for clinical endpoint prediction. In this model, a time-aware transformer is introduced to learn the personalized irregular temporal patterns of medical events, and a hierarchical attention mechanism is deployed to get the accurate patient fusion representation by comprehensively mining the interactions and correlations among multiple types of medical data. We evaluate our model on the MIMIC-III dataset and MIMIC-IV dataset for the task of mortality prediction, and the results show that TERTIAN achieves higher performance than state-of-the-art approaches.

CD-Surv: a contrastive-based model for dynamic survival analysis

Survival analysis, aimed at investigating the relationships between covariates and event time, has exhibited profound effects on health service management. Longitudinal data with sequential patterns, such as electronic health records (EHRs), contain a large volume of patient treatment trajectories, and therefore, provide great potential for survival analysis. However, most existing studies address the survival analysis problem in a static manner, that is, they only utilize a fraction of longitudinal data, ignore the correlations between multiple visits, and usually may not be able to capture the latent representations of patient treatment trajectories. This inevitably deteriorates the performance of the survival analysis. To address this challenge, we propose an end-to-end contrastive-based model CD-Surv to better understand the patient treatment trajectories and dynamically predict the survival probability of a target patient. Specifically, two data augmentation strategies, namely, mask generation and shuffle generation, are adopted to augment the real treatment trajectories documented in the EHR. Based on this, the hidden representations of the real trajectories can be improved by utilizing contrastive learning between augmented and real trajectories. We evaluated our proposed CD-Surv on two real-world datasets, and the experimental results indicated that our proposed model could outperform state-of-the-art baselines on various evaluation metrics.

Disentangled representation for sequential treatment effect estimation

BACKGROUND AND OBJECTIVE

Treatment effect estimation, as a fundamental problem in causal inference, focuses on estimating the outcome difference between different treatments. However, in clinical observational data, some patient covariates (such as gender, age) not only affect the outcomes but also affect the treatment assignment. Such covariates, named as confounders, produce distribution discrepancies between different treatment groups, thereby introducing the selection bias for the estimation of treatment effects. The situation is even more complicated in longitudinal data, because the confounders are time-varying that are subject to patient history and meanwhile affect the future outcomes and treatment assignments. Existing methods mainly work on cross-sectional data obtained at a specific time point, but cannot process the time-varying confounders hidden in the longitudinal data.

METHODS

In this study, we address this problem for the first time by disentangled representation learning, which considers the observational data as consisting of three components, including outcome-specific factors, treatment-specific factors, and time-varying confounders. Based on this, the proposed approach adopts a recurrent neural network-based framework to process sequential information and learn the disentangled representations of the components from longitudinal observational sequences, captures the posterior distributions of latent factors by multi-task learning strategy. Moreover, mutual information-based regularization is adopted to eliminate the time-varying confounders. In this way, the association between patient history and treatment assignment is removed and the estimation can be effectively conducted.

RESULTS

We evaluate our model in a realistic set-up using a model of tumor growth. The proposed model achieves the best performance over benchmark models for both one-step ahead prediction (0.70% vs 0.74% for the-state-of-the-art model, when γ = 3. Measured by normalized root mean square error, the lower the better) and five-step ahead prediction (1.47% vs 1.83%) in most cases. By increasing the effect of confounders, our proposed model always shows superiority against the state-of-the-art model. In addition, we adopted T-SNE to visualize the disentangled representations and present the effectiveness of disentanglement explicitly and intuitively.

CONCLUSIONS

The experimental results indicate the powerful capacity of our model in learning disentangled representations from longitudinal observational data and dealing with the time-varying confounders, and demonstrate the surpassing performance achieved by our proposed model on dynamic treatment effect estimation.

A systematic review on machine learning and deep learning techniques in cancer survival prediction

Cancer is a disease which is characterised by the unusual and uncontrollable growth of body cells. This usually happens asymptomatically and gets spread to other parts of the body. The major problem in treating cancer is that its progress is not monitored once it is diagnosed. The progress or the prognosis can be done through survival analysis. The survival analysis is the branch of statistics that deals in predicting the time of event of occurrence. In the case of cancer prognosis the event is the survival time of the patient from the onset of the disease or it can be the recurrence of the disease after undergoing a treatment. This study aims to bring out the machine learning and deep learning models involved in providing the prognosis to the cancer patients.

Application and potential of artificial intelligence in neonatal medicine

Neonatal care is becoming increasingly complex with large amounts of rich, routinely recorded physiological, diagnostic and outcome data. Artificial intelligence (AI) has the potential to harness this vast quantity and range of information and become a powerful tool to support clinical decision making, personalised care, precise prognostics, and enhance patient safety. Current AI approaches in neonatal medicine include tools for disease prediction and risk stratification, neurological diagnostic support and novel image recognition technologies. Key to the integration of AI in neonatal medicine is the understanding of its limitations and a standardised critical appraisal of AI tools. Barriers and challenges to this include the quality of datasets used, performance assessment, and appropriate external validation and clinical impact studies. Improving digital literacy amongst healthcare professionals and cross-disciplinary collaborations are needed to harness the full potential of AI to help take the next significant steps in improving neonatal outcomes for high-risk infants.

Development and validation of a novel nomogram for predicting the occurrence of myopia in schoolchildren: A prospective cohort study

PURPOSE

Myopia is a major public health issue and occurs at young ages. Apart from its high prevalence, myopia results in high costs and irreversible blinding diseases. Accurate prediction of the risk of myopia onset is crucial for its precise prevention. We aimed to develop and validate an effective nomogram for predicting myopia onset in schoolchildren.

DESIGN

School-based prospective cohort study.

METHODS

A total of 1073 schoolchildren were enrolled from November 2014 to May 2019 in China, and were divided into the training and validation cohorts. Myopia was defined as a spherical equivalent refraction (SER) ≤-0.5 diopters. Predictors of myopia were determined through the least absolute shrinkage and selection operator regression and multivariable Cox proportional hazard model based on the training cohort. The predictive performance of the nomogram was validated internally through time-dependent receiver operating characteristic (ROC) curves, calibration plot, decision curve analysis, and Kaplan-Meier curves.

RESULTS

Independent predictors at baseline including gender, SER, axial length, corneal refractive power, and positive relative accommodation were included in the nomogram prediction model. This nomogram demonstrated excellent calibration, clinical net benefit, and discrimination, with all the area under the ROC curves (AUCs) between 0.74 and 0.86 in the training and validation cohorts. The Kaplan-Meier curves showed that 3 distinct risk groups stratified through X-tile analysis were well discriminated and robust among subgroups. The Harrell's C-index and net reclassification improvement demonstrated that the nomogram substantially improved compared with previous models. An online myopia risk calculator was generated for better individual prediction.

CONCLUTIONS

The nomogram provides accurate and individual prediction of myopia onset in schoolchildren. External validation is needed to verify the generalizability of this nomogram.

Neural Networks for Survival Prediction in Medicine Using Prognostic Factors: A Review and Critical Appraisal

Survival analysis deals with the expected duration of time until one or more events of interest occur. Time to the event of interest may be unobserved, a phenomenon commonly known as right censoring, which renders the analysis of these data challenging. Over the years, machine learning algorithms have been developed and adapted to right-censored data. Neural networks have been repeatedly employed to build clinical prediction models in healthcare with a focus on cancer and cardiology. We present the first ever attempt at a large-scale review of survival neural networks (SNNs) with prognostic factors for clinical prediction in medicine. This work provides a comprehensive understanding of the literature (24 studies from 1990 to August 2021, global search in PubMed). Relevant manuscripts are classified as methodological/technical (novel methodology or new theoretical model; 13 studies) or applications (11 studies). We investigate how researchers have used neural networks to fit survival data for prediction. There are two methodological trends: either time is added as part of the input features and a single output node is specified, or multiple output nodes are defined for each time interval. A critical appraisal of model aspects that should be designed and reported more carefully is performed. We identify key characteristics of prediction models (i.e., number of patients/predictors, evaluation measures, calibration), and compare ANN's predictive performance to the Cox proportional hazards model. The median sample size is 920 patients, and the median number of predictors is 7. Major findings include poor reporting (e.g., regarding missing data, hyperparameters) as well as inaccurate model development/validation. Calibration is neglected in more than half of the studies. Cox models are not developed to their full potential and claims for the performance of SNNs are exaggerated. Light is shed on the current state of art of SNNs in medicine with prognostic factors. Recommendations are made for the reporting of clinical prediction models. Limitations are discussed, and future directions are proposed for researchers who seek to develop existing methodology.

Dynamic Digital Twin: Diagnosis, Treatment, Prediction, and Prevention of Disease During the Life Course

A digital twin (DT), originally defined as a virtual representation of a physical asset, system, or process, is a new concept in health care. A DT in health care is not a single technology but a domain-adapted multimodal modeling approach incorporating the acquisition, management, analysis, prediction, and interpretation of data, aiming to improve medical decision-making. However, there are many challenges and barriers that must be overcome before a DT can be used in health care. In this viewpoint paper, we build on the current literature, address these challenges, and describe a dynamic DT in health care for optimizing individual patient health care journeys, specifically for women at risk for cardiovascular complications in the preconception and pregnancy periods and across the life course. We describe how we can commit multiple domains to developing this DT. With our cross-domain definition of the DT, we aim to define future goals, trade-offs, and methods that will guide the development of the dynamic DT and implementation strategies in health care.

Time-resolved trajectory of glucose lowering medications and cardiovascular outcomes in type 2 diabetes: a recurrent neural network analysis

Aim

Treatment algorithms define lines of glucose lowering medications (GLM) for the management of type 2 diabetes (T2D), but whether therapeutic trajectories are associated with major adverse cardiovascular events (MACE) is unclear. We explored whether the temporal resolution of GLM usage discriminates patients who experienced a 4P-MACE (heart failure, myocardial infarction, stroke, death for all causes).

Methods

We used an administrative database (Veneto region, North-East Italy, 2011–2018) and implemented recurrent neural networks (RNN) with outcome-specific attention maps. The model input included age, sex, diabetes duration, and a matrix of GLM pattern before the 4P-MACE or censoring. Model output was discrimination, reported as area under receiver characteristic curve (AUROC). Attention maps were produced to show medications whose time-resolved trajectories were the most important for discrimination.

Results

The analysis was conducted on 147,135 patients for training and model selection and on 10,000 patients for validation. Collected data spanned a period of ~ 6 years. The RNN model efficiently discriminated temporal patterns of GLM ending in a 4P-MACE vs. those ending in an event-free censoring with an AUROC of 0.911 (95% C.I. 0.904–0.919). This excellent performance was significantly better than that of other models not incorporating time-resolved GLM trajectories: (i) a logistic regression on the bag-of-words encoding all GLM ever taken by the patient (AUROC 0.754; 95% C.I. 0.743–0.765); (ii) a model including the sequence of GLM without temporal relationships (AUROC 0.749; 95% C.I. 0.737–0.761); (iii) a RNN model with the same construction rules but including a time-inverted or randomised order of GLM. Attention maps identified the time-resolved pattern of most common first-line (metformin), second-line (sulphonylureas) GLM, and insulin (glargine) as those determining discrimination capacity.

Conclusions

The time-resolved pattern of GLM use identified patients with subsequent cardiovascular events better than the mere list or sequence of prescribed GLM. Thus, a patient’s therapeutic trajectory could determine disease outcomes.

Prediction of survival in oropharyngeal squamous cell carcinoma using machine learning algorithms: A study based on the surveillance, epidemiology, and end results database

Background

We determined appropriate survival prediction machine learning models for patients with oropharyngeal squamous cell carcinoma (OPSCC) using the “Surveillance, Epidemiology, and End Results” (SEER) database.

Methods

In total, 4039 patients diagnosed with OPSCC between 2004 and 2016 were enrolled in this study. In particular, 13 variables were selected and analyzed: age, sex, tumor grade, tumor size, neck dissection, radiation therapy, cancer directed surgery, chemotherapy, T stage, N stage, M stage, clinical stage, and human papillomavirus (HPV) status. The T-, N-, and clinical staging were reconstructed based on the American Joint Committee on Cancer (AJCC) Staging Manual, 8th Edition. The patients were randomly assigned to a development or test dataset at a 7:3 ratio. The extremely randomized survival tree (EST), conditional survival forest (CSF), and DeepSurv models were used to predict the overall and disease-specific survival in patients with OPSCC. A 10-fold cross-validation on a development dataset was used to build the training and internal validation data for all models. We evaluated the predictive performance of each model using test datasets.

Results

A higher c-index value and lower integrated Brier score (IBS), root mean square error (RMSE), and mean absolute error (MAE) indicate a better performance from a machine learning model. The C-index was the highest for the DeepSurv model (0.77). The IBS was also the lowest in the DeepSurv model (0.08). However, the RMSE and RAE were the lowest for the CSF model.

Conclusions

We demonstrated various machine-learning-based survival prediction models. The CSF model showed a better performance in predicting the survival of patients with OPSCC in terms of the RMSE and RAE. In this context, machine learning models based on personalized survival predictions can be used to stratify various complex risk factors. This could help in designing personalized treatments and predicting prognoses for patients.

Developing machine learning algorithms for dynamic estimation of progression during active surveillance for prostate cancer

Active Surveillance (AS) for prostate cancer is a management option that continually monitors early disease and considers intervention if progression occurs. A robust method to incorporate “live” updates of progression risk during follow-up has hitherto been lacking. To address this, we developed a deep learning-based individualised longitudinal survival model using Dynamic-DeepHit-Lite (DDHL) that learns data-driven distribution of time-to-event outcomes. Further refining outputs, we used a reinforcement learning approach (Actor-Critic) for temporal predictive clustering (AC-TPC) to discover groups with similar time-to-event outcomes to support clinical utility. We applied these methods to data from 585 men on AS with longitudinal and comprehensive follow-up (median 4.4 years). Time-dependent C-indices and Brier scores were calculated and compared to Cox regression and landmarking methods. Both Cox and DDHL models including only baseline variables showed comparable C-indices but the DDHL model performance improved with additional follow-up data. With 3 years of data collection and 3 years follow-up the DDHL model had a C-index of 0.79 (±0.11) compared to 0.70 (±0.15) for landmarking Cox and 0.67 (±0.09) for baseline Cox only. Model calibration was good across all models tested. The AC-TPC method further discovered 4 distinct outcome-related temporal clusters with distinct progression trajectories. Those in the lowest risk cluster had negligible progression risk while those in the highest cluster had a 50% risk of progression by 5 years. In summary, we report a novel machine learning approach to inform personalised follow-up during active surveillance which improves predictive power with increasing data input over time.

Causal inference for observational longitudinal studies using deep survival models

OBJECTIVE

Causal inference for observational longitudinal studies often requires the accurate estimation of treatment effects on time-to-event outcomes in the presence of time-dependent patient history and time-dependent covariates.

MATERIALS AND METHODS

To tackle this longitudinal treatment effect estimation problem, we have developed a time-variant causal survival (TCS) model that uses the potential outcomes framework with an ensemble of recurrent subnetworks to estimate the difference in survival probabilities and its confidence interval over time as a function of time-dependent covariates and treatments.

RESULTS

Using simulated survival datasets, the TCS model showed good causal effect estimation performance across scenarios of varying sample dimensions, event rates, confounding and overlapping. However, increasing the sample size was not effective in alleviating the adverse impact of a high level of confounding. In a large clinical cohort study, TCS identified the expected conditional average treatment effect and detected individual treatment effect heterogeneity over time. TCS provides an efficient way to estimate and update individualized treatment effects over time, in order to improve clinical decisions.

DISCUSSION

The use of a propensity score layer and potential outcome subnetworks helps correcting for selection bias. However, the proposed model is limited in its ability to correct the bias from unmeasured confounding, and more extensive testing of TCS under extreme scenarios such as low overlapping and the presence of unmeasured confounders is desired and left for future work.

CONCLUSION

TCS fills the gap in causal inference using deep learning techniques in survival analysis. It considers time-varying confounders and treatment options. Its treatment effect estimation can be easily compared with the conventional literature, which uses relative measures of treatment effect. We expect TCS will be particularly useful for identifying and quantifying treatment effect heterogeneity over time under the ever complex observational health care environment.

Machine Learning–Based Overall Survival Prediction of Elderly Patients With Multiple Myeloma From Multicentre Real-Life Data

Objective

To use machine learning methods to explore overall survival (OS)-related prognostic factors in elderly multiple myeloma (MM) patients.

Methods

Data were cleaned and imputed using simple imputation methods. Two data resampling methods were implemented to facilitate model building and cross validation. Four algorithms including the cox proportional hazards model (CPH); DeepSurv; DeepHit; and the random survival forest (RSF) were applied to incorporate 30 parameters, such as baseline data, genetic abnormalities and treatment options, to construct a prognostic model for OS prediction in 338 elderly MM patients (>65 years old) from four hospitals in Beijing. The C-index and the integrated Brier score (IBwere used to evaluate model performances.

Results

The 30 variables incorporated in the models comprised MM baseline data, induction treatment data and maintenance therapy data. The variable importance test showed that the OS predictions were largely affected by the maintenance schema variable. Visualizing the survival curves by maintenance schema, we realized that the immunomodulator group had the best survival rate. C-indexes of 0.769, 0.780, 0.785, 0.798 and IBS score of 0.142, 0.112, 0.108, 0.099 were obtained from the CPH model, DeepSurv, DeepHit, and the RSF model respectively. The RSF model yield best scores from the fivefold cross-validation, and the results showed that different data resampling methods did affect our model results.

Conclusion

We established an OS model for elderly MM patients without genomic data based on 30 characteristics and treatment data by machine learning.

Prediction and Risk Stratification of Cardiovascular Disease in Diabetic Kidney Disease Patients

Background

Diabetic kidney disease (DKD) patients are facing an extremely high risk of cardiovascular disease (CVD), which is a major cause of death for DKD patients. We aimed to build a deep learning model to predict CVD risk among DKD patients and perform risk stratifying, which could help them perform early intervention and improve personal health management.

Methods

A retrospective cohort study was conducted to assess the risk of the occurrence of composite cardiovascular disease, which includes coronary heart disease, cerebrovascular diseases, congestive heart failure, and peripheral artery disease, in DKD patients. A least absolute shrinkage and selection operator (LASSO) regression was used to perform the variable selection. A deep learning-based survival model called DeepSurv, based on a feed-forward neural network was developed to predict CVD risk among DKD patients. We compared the model performance with the conventional Cox proportional hazards (CPH) model and the Random survival forest (RSF) model using the concordance index (C-index), the area under the curve (AUC), and integrated Brier scores (IBS).

Results

We recruited 890 patients diagnosed with DKD in this retrospective study. During a median follow-up of 10.4 months, there are 289 patients who sustained a subsequent CVD. Seven variables, including age, high density lipoprotein (HDL), hemoglobin (Hb), systolic blood pressure (SBP), smoking status, 24 h urinary protein excretion, and total cholesterol (TC), chosen by LASSO regression were used to develop the predictive model. The DeepSurv model showed the best performance, achieved a C-index of 0.767(95% confidence intervals [CI]: 0.717–0.817), AUC of 0.780(95%CI: 0.721–0.839), and IBS of 0.067 in the validation set. Then we used the cut-off value determined by ROC (receiver operating characteristic) curve to divide the patients into different risk groups. Moreover, the DeepSurv model was also applied to develop an online calculation tool for patients to conduct risk monitoring.

Conclusion

A deep-learning-based predictive model using seven clinical variables can effectively predict CVD risk among DKD patients and perform risk stratification. An online calculator allows its easy implementation.

A novel deep learning prognostic system improves survival predictions for stage III non-small cell lung cancer

Background

Accurate prognostic prediction plays a crucial role in the clinical setting. However, the TNM staging system fails to provide satisfactory individual survival prediction for stage III non‐small cell lung cancer (NSCLC). The performance of the deep learning network for survival prediction in stage III NSCLC has not been explored.

Objectives

This study aimed to develop a deep learning‐based prognostic system that could achieve better predictive performance than the existing staging system for stage III NSCLC.

Methods

In this study, a deep survival learning model (DSLM) for stage III NSCLC was developed based on the Surveillance, Epidemiology, and End Results (SEER) database and was independently tested with another external cohort from our institute. DSLM was compared with the Cox proportional hazard (CPH) and random survival forest (RSF) models. A new prognostic system for stage III NSCLC was also proposed based on the established deep learning model.

Results

The study included 16,613 patients with stage III NSCLC from the SEER database. DSLM showed the best performance in survival prediction, with a C‐index of 0.725 in the validation set, followed by RSF (0.688) and CPH (0.683). DSLM also showed C‐indices of 0.719 and 0.665 in the internal and real‐world external testing datasets, respectively. In addition, the new prognostic system based on DSLM (AUROC = 0.744) showed better performance than the TNM staging system (AUROC = 0.561).

Conclusion

In this study, a new, integrated deep learning‐based prognostic model was developed and evaluated for stage III NSCLC. This novel approach may be valuable in improving patient stratification and potentially provide meaningful prognostic information that contributes to personalized therapy.

Deep-CSA: Deep Contrastive Learning for Dynamic Survival Analysis with Competing Risks

Survival analysis (SA) is widely used to analyze data in which the time until the event is of interest. Conventional SA techniques assume a specific form for viewing the distribution of survival time as the hitting time of a stochastic process, and explicitly model the relationship between covariates and the distribution of the events hitting time. Although valuable, existing SA models seldom consider to model the dynamic correlations between covariates and more than one event of interest (i.e., competing risks) in the disease progression of subjects. To alleviate this critical problem, we propose a novel deep contrastive learning model to obtain a deep understanding of disease progression of subjects with competing risks from their longitudinal observational data. Specifically, we design a self-supervised objective for learning dynamic representations of subjects suffering from multiple competing risks, such that the relationship between covariates and each specific competing risk changes over time can be well captured. Experiments on two open-source clinical dataset-s, i.e., MIMIC-III and EICU, demonstrate the effectiveness of our proposed model, with remarkable improvements over the state-of-the-art SA models.

Fault Detection and Diagnosis in Industrial Processes with Variational Autoencoder: A Comprehensive Study

This work considers industrial process monitoring using a variational autoencoder (VAE). As a powerful deep generative model, the variational autoencoder and its variants have become popular for process monitoring. However, its monitoring ability, especially its fault diagnosis ability, has not been well investigated. In this paper, the process modeling and monitoring capabilities of several VAE variants are comprehensively studied. First, fault detection schemes are defined in three distinct ways, considering latent, residual, and the combined domains. Afterwards, to conduct the fault diagnosis, we first define the deep contribution plot, and then a deep reconstruction-based contribution diagram is proposed for deep domains under the fault propagation mechanism. In a case study, the performance of the process monitoring capability of four deep VAE models, namely, the static VAE model, the dynamic VAE model, and the recurrent VAE models (LSTM-VAE and GRU-VAE), has been comparatively evaluated on the industrial benchmark Tennessee Eastman process. Results show that recurrent VAEs with a deep reconstruction-based diagnosis mechanism are recommended for industrial process monitoring tasks.

Establishment of a Predictive Model for GvHD-free, Relapse-free Survival after Allogeneic HSCT using Ensemble Learning

Stacked ensemble of machine-learning algorithms could establish more accurate prediction model for survival analysis than existing methods.

Stacked ensemble model can be applied to personalized prediction of HSCT outcomes from pretransplant characteristics.

Graft-versus-host disease-free, relapse-free survival (GRFS) is a useful composite end point that measures survival without relapse or significant morbidity after allogeneic hematopoietic stem cell transplantation (allo-HSCT). We aimed to develop a novel analytical method that appropriately handles right-censored data and competing risks to understand the risk for GRFS and each component of GRFS. This study was a retrospective data-mining study on a cohort of 2207 adult patients who underwent their first allo-HSCT within the Kyoto Stem Cell Transplantation Group, a multi-institutional joint research group of 17 transplantation centers in Japan. The primary end point was GRFS. A stacked ensemble of Cox Proportional Hazard (Cox-PH) regression and 7 machine-learning algorithms was applied to develop a prediction model. The median age for the patients was 48 years. For GRFS, the stacked ensemble model achieved better predictive accuracy evaluated by C-index than other state-of-the-art competing risk models (ensemble model: 0.670; Cox-PH: 0.668; Random Survival Forest: 0.660; Dynamic DeepHit: 0.646). The probability of GRFS after 2 years was 30.54% for the high-risk group and 40.69% for the low-risk group (hazard ratio compared with the low-risk group: 2.127; 95% CI, 1.19-3.80). We developed a novel predictive model for survival analysis that showed superior risk stratification to existing methods using a stacked ensemble of multiple machine-learning algorithms.