TFN Temporal Fusion Nexus: A Multimodal Temporal Fusion Prediction System for Kidney Transplant Patients

TFN is a multimodal deep learning framework for kidney transplant patients. It integrates irregular time-series vital signs, static attributes, and clinical text notes to achieve multi-time-span prediction of kidney transplant rejection, graft loss, and mortality risks. It aims to address the limitations of traditional prognosis assessment that relies on experience and single indicators, and integrate multi-source heterogeneous data to improve prediction accuracy.

Kidney transplantation is the most effective treatment for end-stage renal disease, but long-term management faces challenges. Traditional risk assessment relies on experience and single laboratory indicators, making it difficult to utilize the large amount of heterogeneous data from follow-ups. Modern follow-ups generate three core types of data: 1. Irregular time series (vital signs, laboratory results, medication records with varying sampling frequencies, high missing rates, and large timeline differences); 2. Static attributes (baseline characteristics of donors and recipients, such as age, HLA matching, etc.); 3. Free-text clinical notes (unstructured, containing rich insights but difficult to utilize). TFN is a solution designed to address these challenges.

The core innovation of TFN is the unified encoding and deep fusion of three heterogeneous data types: 1. Time-series encoding: Uses TimeAwareLSTM to explicitly model sampling intervals, distinguishing between missing and unsampled data; optional temporal self-attention to learn time-step dependencies; feature-level missing value handling (marked with value_mask). 2. Static feature fusion: StaticEncoder encodes categorical and numerical features separately before fusion, injects them into temporal hidden states, and captures interactions between baseline risks and dynamic changes (e.g., the significance of creatinine increase in elderly patients). 3. Clinical note encoding: Uses the gte-large model to embed text, interacts with temporal states via cross-attention, and solves the problem of note time alignment.

TFN provides multiple prediction heads: MultiModal head (deterministic prediction, outputting risk probabilities), MultiModalVAE head (generative modeling, capturing uncertainty), and SimpleMLP head (lightweight classifier). It supports joint prediction across multiple time spans (30 days, 90 days, 1 year, etc.). The multi-task design improves data efficiency and learns correlations between different prognostic indicators.

TFN was developed and validated based on the NephroCAGE cohort, providing a complete preprocessing pipeline: 1. Data loading and cleaning (handling naming, units, and outliers); 2. Time-series alignment (unifying time axes and calculating derived indicators); 3. Patient-level partitioning (to avoid data leakage); 4. PyTorch dataset encapsulation (supporting batch processing of variable-length sequences).

TFN includes rich evaluation tools: Performance evaluation (AUROC metric to assess discriminative ability across different time spans); cluster analysis (identifying similar risk subgroups); feature importance (SHAP method); visualization tools (temporal attention weights, risk curves, etc.).

The value of TFN for kidney transplant management includes: Early warning (pre-symptomatic identification of high risks); personalized monitoring (optimizing follow-up frequency); decision support (assisting with biopsies and immune regimen adjustments); and research tool (open-source code to drive progress in the field).

Current limitations: Based on the non-public NephroCAGE cohort, which limits direct reproducibility, but the code framework can be applied to other cohorts. Future directions: Integrating genomic data, introducing causal inference, and developing real-time prediction systems integrated into electronic health records.