Deep Learning Survival Prediction

deep-learning-survival-prediction

Summary

Fit a Cox-nnet deep-learning survival model on metabolomic or expression data paired with event/time vectors to generate personalized risk stratification and prognosis indices. This skill integrates regularization, cross-validation, and gradient-based optimization to produce both individual-level risk scores and pathway-level survival predictions.

When to use

You have metabolomic or expression feature data, sample-level event indicators (e.g., censoring status), and follow-up times, and you want to predict survival risk at both the metabolite/gene and pathway levels while accounting for non-linear relationships between features and survival outcomes.

When NOT to use

• Event/time vectors are missing or incomplete; Cox models require complete follow-up data.
• Feature matrix contains only a handful of metabolites and no pathway grouping is available; deep learning requires sufficient feature dimensionality and sample size to avoid overfitting.
• Survival data violates the proportional hazards assumption; traditional Cox-PH may be more appropriate.

Inputs

• event vector (binary: 0=censored, 1=event occurred)
• time vector (follow-up duration per sample)
• expression/metabolomic feature matrix (samples × features)
• cvlambda parameter (regularization sequence; optional)
• percent parameter (train/test split ratio; optional)

Outputs

• prognosis index (numeric risk scores per sample)
• risk group assignments (categorical, e.g., low/medium/high)
• metabolite-level survival predictions
• pathway-level survival predictions
• cross-validation performance metrics

How to apply

Load the survival event vector, follow-up time vector, and expression/metabolite feature matrix into R. Invoke lilikoi.prognosis() with coxnnet=TRUE, setting alpha=0 (no regularization penalty) or a non-zero value if feature selection is desired, nfold=5 for 5-fold cross-validation, method='quantile' for risk group stratification, and coxnnet_method='gradient' for optimization. The function applies Cox-nnet (a deep-learning variant of the Cox proportional hazards model) with gradient-based fitting to learn non-linear interactions between features and survival. Extract the resulting prognosis index (risk scores per sample) and risk group assignments; these represent personalized survival predictions that can be used for patient stratification or prognostic validation.

Related tools

• Lilikoi v2.0 (R package implementing lilikoi.prognosis() function with Cox-nnet and Cox-PH modules for survival prediction) — https://github.com/lanagarmire/lilikoi2
• Cox-nnet (Deep-learning survival model variant that learns non-linear associations between expression/metabolite features and survival outcomes)
• Cox-PH (Classical Cox proportional hazards baseline model; integrated as alternative to Cox-nnet)
• R (Computational environment and language for executing Lilikoi prognosis workflows)

Examples

lilikoi.prognosis(event=event, time=time, exprdata=exprdata, alpha=0, nfold=5, method="quantile", coxnnet=TRUE, coxnnet_method="gradient")

Evaluation signals

• Prognosis index values are numeric and sample counts match input expression matrix row count.
• Risk group assignments partition samples into categorical strata; verify median survival times differ significantly between groups via log-rank test.
• Cross-validation fold performance metrics (e.g., concordance index) are ≥ 0.6, indicating discrimination better than random chance.
• Prognosis index correlates positively with event occurrence and inversely with follow-up time in validation subset (Spearman ρ > 0.3).
• Pathway-level predictions are derived and distinct from metabolite-level predictions, confirming hierarchical integration.

Limitations

• Cox-nnet assumes proportional hazards; violations may reduce model calibration.
• Requires sufficient sample size and event count (rule of thumb: ≥10–20 events per feature) to avoid overfitting; small cohorts benefit from stronger regularization (alpha > 0).
• Gradient-based optimization may converge to local minima; multiple random seeds or ensemble approaches can mitigate.
• No changelog found in repository documentation; users should verify version compatibility.
• Metabolite-to-pathway mapping depends on MetaTOpathway conversion and HMDB database completeness; missing or ambiguous identifiers may exclude features from pathway analysis.

Evidence

• [other] The lilikoi.prognosis function integrates Cox-nnet with hyperparameters: "The lilikoi.prognosis function accepts survival event and time vectors, expression data, regularization parameter alpha (0 for no penalty), cross-validation fold count (nfold=5), prognosis index"
• [intro] Cox-nnet is a deep-learning survival model: "the most significant addition of prognosis prediction, implemented by Cox-PH model and the deep-learning based Cox-nnet model"
• [intro] Lilikoi v2.0 adds deep-learning classification: "The new Lilikoi v2.0 R package has implemented a deep-learning method for classification, in addition to popular machine learning methods."
• [other] Workflow for invoking lilikoi.prognosis: "Invoke the Lilikoi v2.0 prognosis module with parameters: alpha=0 (no regularization penalty), nfold=5 (5-fold cross-validation), quantile method for risk stratification, and coxnnet=TRUE to fit the"
• [readme] Lilikoi v2 is a comprehensive metabolomics package in R: "Lilikoi v2 is a modern, comprehensive package to enable metabolomics analysis in R programming environment."