ChronoSurv: Graph-Based HNC Survival Analysis

• ChronoSurv is a clinical pathway-guided framework for multimodal survival analysis in head and neck cancer that represents patient care as a hierarchical directed graph.
• It leverages heterogeneous graph attention layers to aggregate information across clinical steps, handling varied modalities and right-censored data.
• The framework achieves state-of-the-art performance on multiple HNC datasets by integrating tabular variables, WSIs, and free-text clinical reports.

Searching arXiv for the primary ChronoSurv paper and closely related multimodal survival-analysis work to ground the article. {"query": "\"ChronoSurv\" multimodal survival analysis head and neck cancer", "max_results": 10} Searching for related multimodal survival-analysis baselines and graph-based prognosis papers mentioned alongside ChronoSurv. {"query": "\"multimodal survival\" graph prognosis arXiv head neck cancer", "max_results": 10} ChronoSurv is a clinical pathway-guided framework for multimodal survival analysis in head and neck cancer (HNC). It represents patient care as a heterogeneous hierarchical directed graph aligned with three key clinical steps—patient background, initial cancer diagnosis, and local surgery—and uses graph neural network layers to produce a patient-level embedding for discrete-time survival prediction under right censoring. The framework is designed for settings in which prognosis depends on heterogeneous information sources acquired at different points in the care pathway, including tabular variables, histopathology whole-slide images (WSIs), and free-text surgical reports or other clinical notes. On HANCOCK, TCGA-HNSC, and their union UniHNC, ChronoSurv is reported to achieve state-of-the-art discriminative performance while maintaining statistically reliable calibration (Miccinilli et al., 17 Jun 2026).

1. Clinical problem and motivation

ChronoSurv targets multimodal survival analysis in HNC, where prognostic assessment depends on demographics, staging, blood tests, pathology variables, WSIs from primary tumor and lymph nodes, and free-text reports. The prediction target is a time-to-event outcome such as overall survival, with explicit handling of right censoring. The central modeling claim is that these data are not merely heterogeneous in representation space; they are also structured by the clinical workflow through which they are acquired (Miccinilli et al., 17 Jun 2026).

The framework is motivated by limitations in both classical and modern survival modeling. Classical methods such as Cox proportional hazards, Weibull models, Random Survival Forests, and gradient-boosted survival trees assume a fixed tabular feature vector, linear or shallow non-linear feature interactions, and no explicit modeling of clinical workflow or time-varying structure in data acquisition. Modern deep survival models, including DeepSurv, DeepHit, TransDSA, and transformer-based survival models, improve expressive power, but in the multimodal setting they typically encode each modality separately, fuse all modality embeddings statically, and pass the fused vector to a survival head. In HNC, this is limiting because clinical data are attached to different steps of care, later-stage information such as surgical pathology builds on earlier steps, and some steps or modalities may be absent in some cohorts.

ChronoSurv addresses these issues by imposing a pathway-aware representation. Rather than treating all observations as a single baseline feature vector, it encodes clinical steps and their temporal order as a directed graph, and it uses relation-specific message passing to model asymmetric interactions across modalities and steps. This makes the model progression-aware in the specific sense used by the paper: information is aggregated in a direction consistent with the workflow background $\rightarrow$ diagnosis $\rightarrow$ surgery.

2. Graph representation of the clinical pathway

ChronoSurv constructs, for each patient, a directed graph $\mathcal{G} = (\mathcal{V}, \mathcal{E})$ with three hierarchical levels. At the fine-grained level, there is one node per sub-modality $k \in \mathcal{K}$, with $\mathcal{K}$ partitioned into WSIs, tabular sub-modalities, and textual reports. Raw input $x_k$ is mapped into a shared $d$-dimensional latent space using a frozen base encoder plus a learnable projector $f_k$: tabular inputs are projected directly, WSIs are encoded by the frozen pathology foundation model UNI, and text is encoded by the frozen BioClinicalBERT encoder. The resulting embedding is denoted $h_k \in \mathbb{R}^d$ (Miccinilli et al., 17 Jun 2026).

At the second level are three clinical-step nodes: $s_1$ for patient background, $\rightarrow$0 for initial cancer diagnosis, and $\rightarrow$1 for local surgery. These step nodes are zero-initialized, $\rightarrow$2, and receive incoming directed edges from the sub-modalities naturally acquired at that step. Examples given include clinical demographics $\rightarrow$3, blood lab node $\rightarrow$4, and tumor WSI node $\rightarrow$5. The step nodes are also connected by temporally ordered edges that encode the care pathway, including next-step edges such as $\rightarrow$6 and $\rightarrow$7, with potential skip edges such as $\rightarrow$8.

At the global level is a single patient node $\rightarrow$9, initialized by mean-pooling over available leaf embeddings:

$\mathcal{G} = (\mathcal{V}, \mathcal{E})$0

Directed edges $\mathcal{G} = (\mathcal{V}, \mathcal{E})$1 connect the step nodes to the patient node. The overall graph therefore comprises fine-grained sub-modality nodes, coarse clinical-step nodes, and a global patient node:

$\mathcal{G} = (\mathcal{V}, \mathcal{E})$2

A distinctive feature is that the topology is patient-specific. If a sub-modality is missing, the corresponding leaf node is absent from $\mathcal{G} = (\mathcal{V}, \mathcal{E})$3 and its incident edges are omitted from $\mathcal{G} = (\mathcal{V}, \mathcal{E})$4. The step nodes and patient node always exist, but the incoming evidence differs by patient. Temporal structure is encoded by step indices and directed edges rather than continuous timestamps, so incomplete workflows are represented structurally rather than imputed numerically.

3. Heterogeneous message passing and progression-aware fusion

ChronoSurv uses three heterogeneous graph attention layers to propagate information up the hierarchy. The first layer, $\mathcal{G} = (\mathcal{V}, \mathcal{E})$5, performs leaf $\mathcal{G} = (\mathcal{V}, \mathcal{E})$6 step aggregation. For a clinical step node $\mathcal{G} = (\mathcal{V}, \mathcal{E})$7, sub-modality features are pooled relation-wise:

$\mathcal{G} = (\mathcal{V}, \mathcal{E})$8

This implements hierarchical pooling that is aware of which sub-modalities feed into which clinical step (Miccinilli et al., 17 Jun 2026).

The second layer, $\mathcal{G} = (\mathcal{V}, \mathcal{E})$9, performs temporal step-to-step propagation over the relation set $k \in \mathcal{K}$0, which includes next-step and skip edges. Its update is residual and layer-normalized:

$k \in \mathcal{K}$1

After this step, each clinical-step node encodes both its locally attached sub-modality information and context from earlier or later steps, which the paper describes as progression-aware dependencies.

The third layer, $k \in \mathcal{K}$2, performs step $k \in \mathcal{K}$3 patient aggregation:

$k \in \mathcal{K}$4

The final patient embedding therefore combines mean-pooled leaf information with step-wise representations after temporal propagation.

Heterogeneity is handled through relation-specific parameters in $k \in \mathcal{K}$5, $k \in \mathcal{K}$6, and $k \in \mathcal{K}$7, so different edge types can carry different attention weights. Directionality is also explicit. Ablations reported in the paper show that replacing heterogeneous modeling with a homogeneous graph degrades performance, and making the graph undirected reduces C-index and worsens IBS. The model thus treats relation type and temporal direction as architectural priors rather than optional implementation detail.

4. Discrete-time survival prediction, censoring, and incomplete modalities

ChronoSurv uses a discrete-time survival formulation. With patient covariates encoded in the graph, true event time $k \in \mathcal{K}$8, censoring time $k \in \mathcal{K}$9, observed time $\mathcal{K}$0, and event indicator $\mathcal{K}$1, the model assumes conditional non-informative censoring and discretizes time as

$\mathcal{K}$2

It predicts the discrete-time hazard

$\mathcal{K}$3

with survival function

$\mathcal{K}$4

Given the final patient embedding $\mathcal{K}$5, the hazard head produces bin-wise hazards

$\mathcal{K}$6

where $\mathcal{K}$7 is an MLP mapping $\mathcal{K}$8 and $\mathcal{K}$9 is the elementwise logistic sigmoid (Miccinilli et al., 17 Jun 2026).

Training uses the discrete-time negative log-likelihood over bins up to the observed time, so event cases are encouraged to have low hazard before the event bin and high hazard at the event bin, while censored cases are encouraged to have low hazard throughout the observed follow-up. ChronoSurv also adds an explicit calibration penalty $x_k$0, giving the total objective

$x_k$1

with $x_k$2 in the experiments. Day-level predictions are obtained by mapping each day to its corresponding time bin.

The same graph design also handles missing modalities. If a sub-modality is absent, it is excluded from $x_k$3, its incident edges are excluded from $x_k$4, and $x_k$5 is mean-pooled over the remaining leaf embeddings. The paper contrasts this with baseline strategies that zero-impute missing features and concatenate them. In the combined UniHNC cohort, this matters because HANCOCK includes clinical, blood, pathology, text, and WSIs from primary tumor and lymph nodes, whereas TCGA-HNSC includes clinical and pathology data with primary tumor WSIs only, and no blood data, no text, and no lymph-node WSIs.

5. Datasets, baselines, and empirical findings

ChronoSurv is evaluated on two public HNC datasets and one combined cohort. The experimental protocol uses 5-fold cross-validation with 70/15/15 train/validation/test splits at patient level, shared frozen encoders (UNI for WSIs and BioClinicalBERT for text), 20,000 training iterations, batch size 8, AdamW with learning rate $x_k$6 and weight decay $x_k$7, cosine annealing, embedding dimension $x_k$8, and $x_k$9 time bins. Main evaluation metrics are Uno’s C-index, Integrated Brier Score (IBS), and the D-Calibration test at Type I error $d$0 (Miccinilli et al., 17 Jun 2026).

Dataset	Patients / censoring	Modalities
HANCOCK	763 / 72.1%	Tabular, WSIs from primary tumor and lymph nodes, text
TCGA-HNSC	526 / 57.6%	Tabular, tumor WSIs, no blood, no text, no lymph-node WSIs
UniHNC	1289 / 66.1%	Combined cohort with different modality availability by institution

The baseline set includes Weibull parametric regression, Cox proportional hazards, Random Survival Forests, DeepSurv, DeepHit, TransDSA, MMD, GraphMMP, HFBSurv, MultiSurv, and SurvPCG. All baselines receive concatenated feature vectors from the same frozen encoders with missing sub-modalities zero-imputed; only ChronoSurv uses pathway-guided graph fusion.

On UniHNC, ChronoSurv achieves the best C-index, $d$1, ahead of SurvPCG at $d$2, and the best IBS, $d$3. On HANCOCK, it achieves the highest C-index, $d$4, and the lowest IBS, $d$5. On TCGA-HNSC, its C-index is $d$6, competitive with MMD and SurvPCG at $d$7, while IBS is $d$8, reported as tied-best with CoxPH and RSF, which have lower C-index. In D-Calibration, ChronoSurv passes in all 5 folds, whereas SurvPCG passes only 1/5 folds; Cox, RSF, and MultiSurv also pass 5/5 but with lower discrimination.

The ablation studies are structurally important. Removing step 1 lowers performance; removing step 2 yields the largest performance hit, with UniHNC C-index dropping from $d$9 to $f_k$0; removing step 3 also degrades performance, including UniHNC C-index to $f_k$1. Removing level 2, so that leaf nodes connect directly to the patient without step nodes, causes a severe drop to UniHNC C-index $f_k$2; removing level 3 yields C-index around $f_k$3. Replacing relation-specific modeling with a homogeneous graph, or removing directionality by making edges undirected, also worsens performance. Table 3 further shows that unimodal variants and flat multimodal fusion schemes—mean pooling, MLP concatenation, gated fusion, modality-only fully connected GNN, cross-attention, and self-attention—do not match the hierarchical pathway-guided design; tabular-only can outperform some naive multimodal fusions, which the paper uses to argue that naive fusion can be detrimental.

6. Interpretation, limitations, and name disambiguation

ChronoSurv includes qualitative analyses intended to expose how pathway structure contributes to prognosis modeling. A contribution matrix derived from normalized magnitudes of message-passing contributions indicates that step 1 acts largely as a self-contained source, step 2 aggregates substantial information from step 1, and step 3 integrates information from all preceding steps. At the patient level, step 1 and step 3 contribute most to the global representation. Kaplan–Meier curves stratified by predicted risk show clearly separated survival curves, with a log-rank test reported as highly significant at $f_k$4. The discrete-time survival head also produces full individual survival distributions that can be visualized as patient-specific survival curves (Miccinilli et al., 17 Jun 2026).

The authors identify several limitations. The work focuses on multimodal fusion rather than feature encoders, and the visual and language encoders are frozen; the paper states that fine-tuning these foundations could further improve performance. The empirical scope is limited to HNC and two cohorts, so broader validation across cancer types and healthcare systems is still needed. The graph itself is defined only at three coarse clinical steps—background, diagnosis, and surgery—so more complex workflows involving multiple therapies, longitudinal follow-up, or recurrence are not explicitly represented. Proposed future directions include extending the framework to larger and more diverse cohorts, exploring alternative or fine-tuned visual and language encoders, and scaling to more granular workflows such as radiotherapy, chemotherapy, and follow-up visits.

A recurrent source of confusion is nomenclature. ChronoSurv in the oncologic survival-analysis sense should be distinguished from Chronos, a proposed ultra-deep near-infrared spectroscopic galaxy survey mission concept (Ferreras et al., 2019); AWS Chronos, a pretrained transformer-based framework for time-series forecasting evaluated under limited tuning (Baron et al., 17 Jan 2025); CHRONOS, a proposed sub-Hz cryogenic torsion-bar gravitational-wave observatory (Inoue et al., 10 Mar 2026); and Chrono, a verifiable logical clock system for Byzantine P2P networks (Sun et al., 2024). In the biomedical literature represented here, ChronoSurv refers specifically to the clinical pathway-guided graph framework for multimodal survival analysis in HNC.