Improving Survival Prediction of Head and Neck Cancer Patients by Modeling Long-term Symptom Burden

Abstract

Accurate survival prediction in head and neck cancer (HNC) is essential for effective treatment planning and management. Patient-reported outcomes (PROs) provide a complementary source of prognostic information by capturing longitudinal symptom trajectories that reflect patients’ health status and treatment response, but their integration into survival modeling remains challenging due to missing data and limited availability at the desired decision points.

We propose an end-to-end approach that models longitudinal symptom trajectories using a bidirectional long short-term memory (Bi-LSTM) network to forecast late symptom burden from baseline PROs. Predicted symptom trajectories are summarized into cluster-level average symptom burden scores, which are standardized and incorporated as continuous covariates into a Cox proportional hazards model alongside clinical features. The proposed approach improves test-set prognostic discrimination, increasing the concordance index from 0.802 to 0.842. These results demonstrate that forecasting longitudinal PRO dynamics and representing symptom burden as continuous, data-driven measures can enhance personalized survival prediction in HNC.

I. Introduction

Advances in therapeutic approaches have considerably improved the survival of patients with head and neck cancer (HNC). However, many patients continue to experience treatment-related side effects that vary in severity from the acute treatment phase through post-therapy recovery, with certain symptoms persisting well into long-term survivorship, leading to severe symptom burden and adversely affecting quality of life (QoL) [1]–[3].

To better understand and manage persistent symptom burden, clinicians increasingly rely on patient-reported outcomes (PROs) collected outside the confines of traditional clinical encounters [4]. By completing questionnaires, patients provide direct assessments of their symptom severity during and after treatment, empowering clinicians and researchers to identify patterns, guide interventions, and ultimately improve QoL [5]. PROs also carry prognostic value [6]–[8] and can complement clinician-assessed measures such as performance status [10], making them highly relevant for prediction modeling [11].

One validated tool for capturing PROs is the M.D. Anderson Symptom Inventory (MDASI) questionnaire, developed at the M.D. Anderson Cancer Center (MDACC) [12]. Tailored specifically for HNC, the MDASI head and neck module (MDASI-HN) [13] contains 28 questions, including 22 core HNC-symptom-related items. Patients repeatedly complete this questionnaire during and after treatment, providing a structured and validated means to monitor symptom trajectories and to guide clinical decision-making over time.

Incorporating PROs into predictive models is important for enabling more personalized and effective cancer care. However, such integration remains challenging due to irregular collection intervals, missing or incomplete data, and the difficulty of translating high-dimensional symptom trajectories into clinically actionable predictors [14]. While prior work has demonstrated the prognostic relevance of PROs [8], [9], practical methods for robustly integrating longitudinal symptom information into survival models remain limited, particularly when long-term PROs are unavailable at the time of treatment planning.

A limitation of prior approaches is that they rely on patient-level clustering or discrete stratification, where individuals are assigned to symptom burden groups based on population-wide patterns. To address this, we adopt a symptom-centered representation that aggregates symptom severities within longitudinal symptom clusters obtained by exploratory factor analysis (EFA), computed at the patient level. By summarizing cluster-level symptom burden as continuous measures derived from each patient’s own ratings, this approach preserves individual variation while providing an interpretable and transferable representation for survival modeling. Notably, symptom clusters are defined at the symptom level and remain fixed across patients, while burden scores are computed independently for each patient without reference to other individuals in the cohort.

Furthermore, because long-term symptom burden is not observed at the time of diagnosis or treatment planning, when survival models are typically applied, we also forecast long-term symptom burden to enable its integration into prognostic modeling.

Specifically, we first identify longitudinal symptom clusters via EFA over the MDASI symptoms [15] and compute cluster-level average symptom severities based on individual predicted patient ratings. To predict patient-specific symptom trajectories, we apply a bidirectional long-short-term memory (Bi-LSTM) recurrent neural network to learn temporal dependencies across multiple symptoms. Predicted symptom ratings at 12 months post-treatment are aggregated within each symptom cluster to derive continuous long-term symptom burden scores, which are standardized and integrated with clinical variables to predict survival.

The result is an end-to-end framework that integrates longitudinal symptom dynamics, data-driven symptom burden representation, and survival prediction for HNC patients. Using baseline PROs and routinely available clinical and treatment information, the framework enables prediction of long-term symptom burden at clinically relevant decision points. Our results show that incorporating Bi-LSTM-derived symptom-burden scores improves test-set prognostic discrimination for overall survival, increasing the concordance index (c-index) from 0.802 to 0.842.

The main contributions of this work are as follows.

• We identify clinically meaningful longitudinal symptom clusters in HNC using EFA and consensus clustering.

• We propose a patient-specific, cluster-based symptom burden representation defined as continuous summaries of symptom severity per longitudinal symptom cluster.

• We introduce a forecast-then-integrate strategy that enables long-term symptom burden to inform survival prediction at treatment planning by forecasting future PRO trajectories from baseline PROs using a multi-symptom Bi-LSTM model.

• We demonstrate that incorporating predicted symptom burden into a Cox proportional hazards model improves overall survival prediction beyond clinical variables alone, yielding higher c-indices and consistently improved time-dependent area under the ROC curves (AUCs) across all evaluated follow-up horizons.

II. Related Work

Reflecting patients’ self-assessments of their health, PROs have gained traction in predictive modeling to support personalized treatment strategies [16], [17] and better characterize symptom burden after therapy [18]. When combined with clinical indicators, PROs have further demonstrated utility in detecting adverse events that affect the QoL and in guiding clinical decision-making for cancer treatment management [19]. In addition, PROs can help identify correlated symptoms [20], [21].

However, using PROs in predictive models for clinical practice remains challenging, with missing data representing a significant hurdle [14]. To tackle this issue, collaborative filtering (CF) has been employed to impute missing PRO information by leveraging known ratings from other symptoms or from the same patient at different time points [22]. This approach results in a relatively robust representation of the data, yet it cannot accommodate the prediction of future PRO ratings.

Recurrent Neural Networks (RNNs) [23] are widely used for supervised longitudinal prediction because they are inherently designed to model sequential data.

An especially popular type of RNN is the Long Short-Term Memory (LSTM) network [24], which has gained traction for both general longitudinal data prediction [25] and a variety of healthcare applications [26], [27]. Examples of LSTM use in healthcare include mimicking pathologists’ decisions in diagnostic workflows [28], [29], classifying sleep patterns in multivariate time-series clinical data [30], and predicting symptom severity in both acute and late stages of treatment [31], [32]. Building upon LSTMs, Bi-LSTM [33] leverages both past and future inputs to enhance performance further. This extended capacity typically allows Bi-LSTM networks to outperform traditional unidirectional LSTMs [34], making them well-suited for complex, time-dependent prediction tasks.

Several studies have explored strategies for summarizing PROs in ways that facilitate their integration into clinical decision models [35], [36]. Prior approaches include hierarchical clustering of longitudinal PROs to identify symptom severity patterns [8], latent profile analysis to characterize patient subgroups based on symptom trajectories [37], and rule-based definitions of symptom complexity derived directly from PRO ratings [38]. While these methods provide useful abstractions of symptom data, they typically rely on patient-level stratification or predefined severity rules.

In contrast, our work focuses on forecasting longitudinal PRO trajectories and summarizing symptom burden as continuous, patient-specific measures defined over clinically meaningful symptom clusters. By integrating Bi-LSTM-predicted symptom trajectories with cluster-level aggregation and survival modeling, our approach avoids discrete patient assignment while enabling symptom burden to inform prognostic prediction at clinically relevant decision points.

III. Proposed Approach

To integrate long-term symptom burden into survival modeling and evaluate its prognostic relevance, we propose the end-to-end methodological framework illustrated in Figure 1. The framework consists of five main components: longitudinal PROs, symptom clusters, Bi-LSTM-based symptom imputation and prediction, data-driven symptom burden representation, and survival analysis. Each component is described in detail in the following subsections.

Fig. 1.

Proposed end-to-end framework for symptom burden-informed survival prediction. Longitudinal PROs are imputed and forecast using Bi-LSTMs to generate predicted PRO trajectories. In parallel, EFA is applied to observed PROs to identify symptom clusters. Predicted PRO trajectories are aggregated within each cluster to derive continuous, cluster-level symptom burden scores. These symptom burden scores are combined with clinical covariates and incorporated into a Cox proportional hazards model for overall survival prediction.

A. Data

In this work, we use data collected using the MDASI-HN questionnaire [13], a specialized instrument developed by MDACC, from 669 HNC patients who were treated using radiation therapy (RT) between 2010 and 2021. We exclude the patients with absence of answering the entire questionnaire. MDASI-HN captures PROs for 28 symptoms, including 22 cancer-related symptoms shown in Table I and 6 general life-interference ratings (e.g., enjoyment of life, mood, work). Since life-interference ratings reflect broader, subjective functional impacts, this study focuses exclusively on the 22 cancer-related symptoms. Of these, 13 (e.g., pain, fatigue, dry mouth) reflect core cancer issues, while 9 (e.g., mucus, swallowing, taste) are specific to HNC cancer (Table I). Each symptom is scored on a 0–10 scale, where 0 indicates the symptom is not experienced, and 10 denotes the worst imaginable severity.

TABLE I.

The 22 symptoms in the MDASI-HN questionnaire by toxicity type. Core refers to core MDASI cancer symptoms, whereas HNC stands for head and neck-specific symptoms.

Type	Symptoms
Core	pain, fatigue, nausea, sleep, distress, shortness of breath (SOB), memory, appetite, drowsiness, dry mouth, sadness, vomiting, numbness
HNC	mucus, swallowing, choke, voice, skin, constipation, taste, mucositis, teeth

The PROs are collected from HNC patients at their initial baseline (B) screening appointment, and the same questionnaire is subsequently administered at the end of treatment (W0) and again at 6 weeks (W6), 6 months (M6), and 12 months (M12) post-treatment. In addition to PROs, clinical variables including demographics, disease information, and treatment were extracted from the electronic health records (EHRs).

This retrospective study was exempt under M.D. Anderson IRB protocol RCR-003–0800. In compliance with the Health Insurance Portability and Accountability Act (HIPAA), informed consent was waived and approved by the IRB, as all analyses were performed over retrospective anonymized data.

B. Symptom Clusters

To identify longitudinal symptom clusters, we first applied EFA for each time point. Symptom ratings were imputed using linear interpolation and back and forward fill. To reduce the patient variability in the ratings, the 10-point scale was discretized using four severity levels: 0: No present, 1: low score (0–3], 2: Mild to moderate score (3–7], and 3: Severe score (7–10].

Bartlett’s test of sphericity and the Kaiser-Meyer-Olkin (KMO) measure were computed to assess the appropriateness of the data for EFA. A KMO value above 0.9 was considered excellent. The initial factor extraction was performed without rotation to obtain eigenvalues. The number of factors was determined using the Kaiser criterion (eigenvalues > 1). EFA was re-run using varimax rotation. To simplify interpretation, sparse loadings were computed by retaining only the highest loading per symptom. The resulting sparse factor loadings were used to define symptom clusters for the given time point, with each symptom assigned to the factor on which it loaded most strongly.

To derive a robust representation of symptom clusters across time points, we implemented a consensus clustering approach using agglomerative hierarchical clustering over the symptom-factor associations to identify stable and recurring symptom clusters. The hierarchical clustering method utilized Ward’s linkage criterion to minimize intra-cluster variance. The resulting dendrogram was inspected to determine an optimal number of consensus clusters, representing symptom groupings that persisted or re-emerged across multiple time points.

C. Bi-LSTM Imputation and Prediction

We model the longitudinal PRO trajectory using the Bi-LSTM recurrent neural network. The Bi-LSTM architecture is well-suited for sequential data, as it processes temporal information in both forward and backward directions, enabling the model to capture patient-specific symptom dynamics across the course of the treatment more effectively than a standard unidirectional LSTM.

For each patient, PROs are represented as a sequence of length T, where each time step consists of a 22-dimensional vector corresponding to MDASI-HN symptom ratings. To address missing symptom ratings at baseline, the model integrates symptom-to-symptom CF-imputation [22] that uses correlations among symptoms rated by the same patient across all time points. This baseline imputation provides a consistent starting profile for each patient, ensuring that the Bi-LSTM receives complete input at the initial time point.

The Bi-LSTM network consists of one or more bidirectional LSTM layers, with H hidden units in each direction. At each time step, the forward and backward hidden states are concatenated to form a 2H-dimensional latent representation. The final Bi-LSTM output is passed through a fully connected layer that produces a 22-dimensional output vector, corresponding to predicted symptom ratings. The model is trained to simultaneously impute missing post-baseline ratings and predict future symptom ratings by minimizing mean squared error between predicted and observed ratings.

To generate imputed post-baseline symptom ratings without introducing information leakage, we adopt an iterative, outof-fold prediction strategy. Specifically, we train four time-step specific Bi-LSTM models with progressively increasing input lengths, where each model predicts symptom ratings at the next time point using all available prior observations. For each time step, three-fold cross-validation is employed where models are trained on two folds and used to generate predictions for the held-out fold. These out-of-fold predictions are then used to fill missing symptom ratings, yielding a completed symptom trajectory composed entirely of Bi-LSTM predicted values at imputed positions. The accuracy of Bi-LSTM-based symptom imputation and prediction is evaluated on held-out folds using standard error metrics, as described in the Experiments section.

Model parameters and hyperparameters are optimized jointly to maximize predictive performance. Bayesian optimization is applied to efficiently explore the parameter space, including the number of Bi-LSTM layers (1–4), the number of hidden units per layer (10–150), optimizer choice (Adam or SGD), and learning rate ranges (10⁻⁴-10⁻² for Adam, and 10⁻³-1 for SGD). The implementation was carried out in TensorFlow and trained on an NVIDIA RTX 4080 GPU.

As summarized in Table II, Bayesian optimization produced fold-specific optimal hyperparameter configurations for the Bi-LSTM models at each time step. Despite fold-level variability, deeper Bi-LSTM architectures were selected for early-stage models with limited temporal input, while later-stage models consistently favored shallower architectures as longitudinal context increased. Hidden dimensionality and optimizer choice varied across models. These hyperparameter configurations were used to generate out-of-fold imputed and predicted symptom values throughout the study.

TABLE II.

Representative Bi-LSTM Hyperparameter Configurations for Each Time-Step Model Identified by Bayesian Optimization

Model	# Layers	Units	Optimizer	LR
T = B→W0	4	150	adam	0.003
T = W0→W6	2	70	sgd	0.019
T = W6→M6	1	80	adam	0.007
T = M6→M12	1	140	sgd	0.239

D. Data-Driven Symptom Burden Representation

To integrate longitudinal symptom information into survival modeling, it is necessary to summarize high-dimensional PROs into a compact representation that preserves prognostic signal while remaining interpretable. Rather than relying on fixed or externally defined symptom severity thresholds, we represent symptom burden using continuous, data-driven summaries derived directly from patient-reported symptom ratings.

Based on the clinically coherent symptom groupings identified through EFA, the 22 MDASI-HN symptoms are organized into three symptom clusters, denoted as SC_k for k = 1, 2, 3, each representing a distinct pattern of co-occurring symptoms. For each patient and each symptom cluster, we compute the average symptom severity across all symptoms belonging to that cluster. This cluster-level average provides a continuous summary of symptom burden that is specific to the patient and does not involve patient-level stratification or discrete assignment.

At the inference stage, predicted symptom ratings at M12 are obtained from the Bi-LSTM model. For each cluster SC_k, the predicted symptom values are averaged to produce a continuous long-term symptom burden score. To ensure comparability across clusters and covariates, these burden scores are standardized using z-score normalization prior to inclusion in the survival model. The resulting standardized cluster-level symptom burden scores are incorporated as continuous covariates, together with clinical features, in the Cox proportional hazards model for survival analysis.

E. Survival Analysis

To examine the relationship between symptom burden and overall survival, we first conduct univariate survival analysis using Kaplan-Meier (KM) curves. For visualization purposes, patients are stratified within each symptom cluster based on the cohort mean of the average cluster-level symptom ratings, yielding high- and low-burden groups. Differences in survival distributions are evaluated to illustrate unadjusted associations between cluster-specific symptom burden and overall survival.

To quantify the prognostic value of symptom burden relative to clinical features, we fit lasso-regularized Cox proportional hazards models using continuous symptom burden scores. Three models are constructed:

1. a clinical-only model including demographic and treatment variables.

2. a symptom-burden-only model including standardized cluster-level symptom burden scores.

3. a combined model incorporating both clinical features and symptom burden scores.

An L1 penalty with a fixed penalization strength of 10⁻² is applied uniformly across covariates to promote sparsity and mitigate multicollinearity.

For each fitted model, regression coefficients are used to estimate log-hazard ratios for individual covariates. Survival probabilities beyond specified time horizons t, ranging from 12 to 60 months in 6-month increments, are estimated using the fitted Cox models. For each categorical covariate, one reference category is excluded prior to model fitting to ensure identifiability and interpretable hazard ratios.

Model discrimination is evaluated using the c-index and time-dependent AUC across multiple survival horizons, as reported in the Experiments section.

IV. Experimental Results

In this section, we first describe the experimental setup and data statistics. We then present the results of the EFA used to identify longitudinal symptom clusters. Next, we evaluate the accuracy of Bi-LSTM-based symptom imputation and short-horizon prediction compared with baseline methods. We subsequently report descriptive statistics of cluster-level symptom burden and illustrate patient stratification for visualization purposes. Finally, we evaluate survival prediction performance and assess the impact of incorporating continuous symptom burden measures into regularized Cox models.

A. Experimental Setup and Data Statistics

We analyzed clinical and PROs from 669 patients. Different components of the proposed framework employ distinct evaluation strategies. Bi-LSTM-based symptom imputation and prediction are evaluated using three-fold cross-validation with out-of-fold prediction to avoid information leakage. In contrast, survival analysis is evaluated using a stratified 80/20 train/test split based on overall survival status (training: N = 536 and testing: N = 133). All symptom clustering and computation of cluster-level symptom burden summaries are performed using the training data and subsequently applied to the held-out test set.

Nine clinical features were included in the survival analysis: age, induction chemotherapy (IC), concurrent chemotherapy (CC), tumor laterality, smoking status, tumor site, N stage, T stage, and gender. All categorical variables were encoded using one-hot encoding, with one reference category per covariate excluded (marked with * in Table III) to ensure identifiability in the Cox models.

TABLE III.

Patient distribution for clinical features, PRO, and survival

Feature	Category	Train (N=536) N (%)	Test (N=133) N (%)
Age	≥ 65	178 (33.2)	45 (33.8)
	< 65*	358 (66.8)	88 (66.2)
Concurrent	Yes	410 (76.5)	107 (80.5)
	No*	126 (23.5)	26 (19.5)
Tumor Laterality	Left	252 (47.0)	60 (45.1)
	Right	268 (50.0)	72 (54.1)
	Other*	16 (3.0)	1 (0.8)
Smoking Status	Yes	264 (49.3)	65 (48.9)
	No*	272 (50.7)	68 (51.1)
Site of Tumor	BOT	270 (50.4)	61 (45.9)
	Tonsil*	244 (45.5)	67 (50.4)
	Other	22 (4.1)	5 (3.8)
N Stage	N0–1*	397 (74.1)	99 (74.4)
	N2–3	139 (25.9)	34 (25.6)
T Stage	T0–1*	180 (33.6)	49 (36.8)
	T2	200 (37.3)	43 (32.3)
	T3–4	156 (29.1)	41 (30.8)
IC	Yes	111 (20.7)	25 (18.8)
	No*	425 (79.3)	108 (81.2)
Gender	Male	491 (91.6)	120 (90.2)
	Female*	45 (8.4)	13 (9.8)
Available PRO	B	469 (87.5)	110 (82.7)
	W0	345 (64.4)	80 (60.1)
	W6	296 (55.2)	68 (51.1)
	M6	368 (68.7)	84 (63.2)
	M12	299 (55.8)	68 (51.1)
Follow-up	Median (Q1-Q3)	28 (18–53)	27 (18–46)
Overall Survival	Deceased	54 (10.1)	12 (9.0)
	Alive	482 (90.0)	121 (91.0)

^*Reference categories in the Cox model.

Table III summarizes the distribution of the clinical features, PRO response rate, and overall survival for both the training and testing sets used in survival modeling. The two sets show comparable distributions across clinical features, PRO availability, and survival outcomes. In both sets, the majority of patients were male (91.6% / 90.2%), and approximately half were current smokers (49.3% / 48.9%). Most patients presented with early T or N stages, with tumors predominantly located at the base of the tongue or in the tonsils. More than 75% of patients received concurrent chemotherapy, whereas approximately 20% underwent induction chemotherapy. PRO completion rates exceeded 50% across all time points, with baseline surveys showing the highest participation (>80%). Median follow-up time was 28 months in the training set and 27 months in the testing set, with approximately 10% of patients reported to be deceased.

B. Symptom Clusters

Table IV summarizes the EFA metrics for each time point and the number of sparse factors identified (where the eigenvalue was greater than 1). All time points demonstrated strong suitability for EFA, with highly significant Bartlett’s tests and excellent KMO values (> 0.9). The number of extracted factors ranged from 3 to 4 across the time points.

TABLE IV.

Exploratory Factor Analysis Metrics Across Timepoints

Time	Bartlett’s Test (χ2, p)	KMO	Factors*
B	7516.71, p < 0.001	0.934	3
W0	7204.91, p < 0.001	0.910	3
W6	6205.75, p < 0.001	0.922	4
M6	6212.94, p < 0.001	0.918	4
M12	6583.70, p < 0.001	0.927	4

^*Factors are set where Eigenvalue > 1.

Figure 2 presents the sparse loadings for the symptoms, grouped by time point and factor. These loadings highlight consistent symptom groupings and shifts in cluster composition over the course of treatment and recovery. Colors are used to indicate the cluster identified through consensus clustering. As can be seen, vomiting and nausea appear together as the only two symptoms in a factor for all time points. For other symptoms/factors, there are small variations across time points.

Fig. 2.

Sparse Loadings for each time point.

Figure 3 shows the resulting dendrogram after applying consensus clustering using agglomerative hierarchical clustering on the sparse factor loadings derived from exploratory EFA. The three symptom clusters and the corresponding symptoms are identified in Table V.

Fig. 3.

Longitudinal Consensus Symptom Cluster.

TABLE V.

The three symptom clusters identified by consensus clustering.

Cluster	Symptoms
SC 1	fatigue, sleep, distress, sadness, numbness shortness of breath (SOB), memory, drowsiness
SC 2	mucus, swallowing, choke, voice, skin, constipation, taste, mucositis, teeth, dry mouth, pain, appetite
SC 3	nausea, vomiting

C. Evaluation of Bi-LSTM Imputation and Prediction

We evaluate the accuracy of the proposed Bayesian-optimized Bi-LSTM for both imputation and one-step-ahead prediction across the 22 MDASI symptoms at the longitudinal time points. All results were obtained using 3-fold cross-validation, and performance is reported as RMSE and MAE (mean ± SD) across the held-out folds (Table VI).

TABLE VI.

Imputation and Prediction Accuracy of Bi-LSTM

Method	Task	RMSE (mean ± SD)	MAE (mean ± SD)
Linear interpolation	Imputation	2.68 ± 0.06	1.67 ± 0.03
Mean		2.14 ± 0.08	1.56 ± 0.03
Median		2.28 ± 0.12	1.41 ± 0.08
Bi-LSTM (proposed, Bayes)		2.05 ± 0.07	1.46 ± 0.03
Linear interpolation + LOCF	Prediction	2.58 ± 0.04	1.66 ± 0.05
Mean + LOCF		2.76 ± 0.05	1.88 ± 0.04
Median + LOCF		2.77 ± 0.05	1.83 ± 0.05
Bi-LSTM (proposed, Bayes)		2.01 ± 0.02	1.46 ± 0.02

As shown in Table VI, the proposed Bi-LSTM achieved the best overall imputation accuracy, with an RMSE of 2.05 ± 0.07. Although column-wise median imputation achieved a slightly lower MAE (1.41 ± 0.08), Bi-LSTM showed lower RMSE and more stable performance across folds. In particular, relative to linear interpolation (2.68 ± 0.06 RMSE and 1.67 ± 0.03 MAE), Bi-LSTM reduced RMSE by approximately 24% and MAE by approximately 13%, indicating improved recovery of longitudinal symptom trajectories beyond simple smoothness-based assumptions.

For one-step-ahead prediction, Bi-LSTM again achieved the best performance, with an RMSE of 2.01 ± 0.02 and an MAE of 1.46 ± 0.02, outperforming all baseline methods. Compared with the strongest baseline, linear interpolation combined with last observation carried forward (LOCF) (2.58 ± 0.04 RMSE and 1.46 ± 0.02 MAE), Bi-LSTM reduced RMSE by approximately 22% and MAE by approximately 12%. Notably, the low standard deviations across folds indicate consistent performance and good generalization to held-out subsets.

D. Data-Driven Symptom Burden Representation

For survival-related experiments, we use the stratified 80/20 train/test split described in the Experimental Setup. To support visualization of univariate survival differences, we summarize cluster-level symptom burden using average symptom ratings within each cluster and dichotomize patients based on the cohort mean of these averages.

Specifically, for each symptom cluster and each patient, we compute the average symptom severity using Bi-LSTM-predicted symptom ratings at M12. For visualization purposes only, patients are stratified into low- and high-burden groups using the cohort mean of the average cluster-level symptom ratings computed from the training set, and the same thresholds are applied to the held-out test set to avoid information leakage.

Table VII reports the resulting cohort mean and the corresponding distribution of patients classified as low- or high-burden within each symptom cluster for the training and testing sets. As shown, the stratification yields balanced group sizes across clusters and comparable distributions between the training and test sets, supporting stable KM visualization without inducing extreme or highly imbalanced groupings. These stratified groups are used exclusively for KM survival curve visualization, whereas continuous, standardized cluster-level symptom burden scores are used as covariates in all Cox models.

TABLE VII.

Cohort Mean of Average Cluster-Level Symptom Ratings and Patient Distribution at M12 (Visualization Only)

Cluster	Mean	Train: Low/High (%)	Test: Low/High (%)
SC 1	1.187	296/240 (55.2/44.8)	68/65 (51.1/48.9)
SC 2	1.387	280/256 (52.2/47.8)	68/65 (51.1/48.9)
SC 3	0.309	299/237 (55.8/44.2)	70/63 (52.6/47.4)

E. Survival Analysis

Figure 4 presents KM survival curves for the full cohort stratified by the symptom burden indicator (high (1) vs. low (0)) for SC₁, SC₂, and SC₃, where burden status is defined using the cohort mean of the cluster-level average symptom severity at M12. Numbers below each curve indicate patients at risk, censored, and experiencing events at selected time points, illustrating the decreasing risk set and increasing uncertainty towards later follow-up, up to 60 months. Across all three symptom clusters, patients classified as high burden show consistently low survival probabilities compared with low-burden patients, with visible separation between curves emerging early and persisting through long-term follow-up. The separation is most pronounced for SC₁ and SC₂, where log-rank tests indicate statistically significant differences between burden groups (p < 0.01), indicating a robust univariate association between high symptom burden and increased mortality risk.

Fig. 4.

KM survival curves stratifies by high (1) vs. low (0) symptom burden for SC₁ in 4a, SC₂ in 4b, and SC₃ in 4c in the full cohort.

For SC₃, although separation between high- and low-burden groups is observable under mean-based stratification, the difference is comparatively modest and accompanied by greater overlap in confidence intervals at intermediate follow-up times. This suggests that symptom burden in SC₃ may be less robust as an isolated prognostic factor. Importantly, KM curves are presented here for descriptive visualization of univariate associations. In the multivariable survival analysis, cluster-level symptom burden is incorporated as a continuous, standardized covariate alongside clinical features, with lasso regularization determining its retained contribution.

We then fit lasso-regularized Cox proportional hazards models on the training set and evaluate performance on the held-out test set. Three models are considered: Clinical only (clinical covariates only), SB only (symptom burden scores only), and Clinical SB (clinical covariates plus symptom burden scores). Figure 5 shows the estimated log-hazard ratios (95% CI) for the clinical SB model. Consistent with established clinical knowledge, advanced T stage (T3–4) (log(HR) = 0.544) and N stage (N2–3) (log(HR) = 0.319) are associated with increased mortality risk.

Fig. 5.

Log-hazard ratios (95% CIs) from the lasso-regularized Cox model using clinical features and cluster-level symptom burden scores.

Among the symptom burden scores, only the SC₂ burden is retained with a clearly positive coefficient (log(HR) = 0.300), while SC₃ only preserves a small positive coefficient (log(HR) = 0.052), and SC₁ is fully shrunk toward zero, suggesting limited incremental contribution after adjustment for clinical covariates. Although the 95% confidence interval for the SC₂ coefficient overlaps zero, similar uncertainty is observed across several established clinical predictors, reflecting the limited number of observed events and the use of L1 regularization.

Predictive discrimination is evaluated on the held-out test set using the c-index and time-dependent AUC at time horizons from 12 to 60 months in 6-month increments (Figure 6). Incorporating standardized cluster-level symptom burden scores with clinical covariates improves discrimination relative to the clinical-only model (c-index: 0.842 vs. 0.802), with consistent gains in time-independent AUC across evaluated time horizons. These improvements indicate that symptom burden information, particularly that retained under regularization, contributes complementary prognostic value beyond standard clinical factors.

Fig. 6.

Time-dependent AUC on the held-out test set for Clinical only, SB only, and Clinical SB Cox models at time cutoffs from 12 to 60 months.

V. Discussion

The MDASI-HN instrument captures a broad set of 22 symptoms that are often interdependent [15] and evolve longitudinally throughout and after treatment. Effectively leveraging this information for prognostic modeling is challenging due to missing data, irregular follow-up, and the high dimensionality of symptom trajectories. In this work, we address these challenges by combining CF-based baseline imputation with iterative Bi-LSTM modeling, enabling recovery and forecasting of patient-specific symptom trajectories even when observed PROs are sparse or unavailable. This substantially expands the usable dataset and allows symptom information to be incorporated at clinically relevant decision points, such as treatment planning.

Using EFA and consensus clustering, we identified three longitudinal symptom clusters consistent with prior reports of symptom co-occurrence in HNC [15]. Rather than relying on fixed, clinician-defined severity thresholds or cohort-level patient stratification, we summarize symptom burden at the cluster level by computing average cluster-specific symptom severity. This representation preserves the underlying symptom structure while providing an interpretable, low-dimensional summary that can be directly incorporated into survival models. This approach avoids ad-hoc severity cutoffs and provides a principled bridge between high-dimensional PRO trajectories and downstream prognostic modeling.

Among the identified clusters, only the symptom burden associated with SC₂, which includes taste disturbance, xerostomia, and swallowing difficulty, demonstrated consistent prognostic relevance. These symptoms are well recognized as persistent, treatment-related toxicities in HNC survivors [40]–[42] and have been linked to nutritional compromise, functional impairment, and long-term QoL deficits [43], [44]. In contrast, the burden score for SC₁ (predominantly fatigue and neuropsychological symptoms) and SC₃ (nausea and vomiting) did not provide stable survival discrimination after adjustment for clinical covariates. This finding suggests that not all symptom domains contribute equally to mortality risk, and highlights the importance of data-driven selection of symptom representations rather than assuming uniform prognostic relevance across symptom types.

Importantly, although individual symptom burden coefficients did not always reach conventional statistical significance under lasso regularization, incorporating predicted long-term symptom burden alongside clinical features consistently improved out-of-sample discrimination. This observation underscores the value of PRO-derived features at the model level, even in settings with limited event rates, and supports the use of predictive performance metrics in conjunction with traditional hazard ratio interpretation.

Several limitations should be noted. First, this study is based on data from a single institution, which may limit generalizability. External validation in multi-institutional cohorts will be necessary to confirm the robustness of the identified symptom clusters and burden representations. Second, symptom burden was summarized using cluster-level averages as alternative representations [45], including symptom-specific weighting schemes or burden indicators, are worth further investigation. Prior work [46] has suggested that symptom severity thresholds may vary across domains, and incorporating such heterogeneity represents an important direction for future research.

Despite these limitations, this study demonstrates that forecasting longitudinal PRO trajectories and translating them into data-driven symptom burden representations can meaningfully enhance survival modeling. By integrating symptom dynamics with clinical features in an end-to-end framework, our approach offers a scalable pathway toward more personalized prognostication and highlights the potential role of PRO-informed models in guiding long-term survivorship care.

VI. Conclusion

In this study, we present an end-to-end framework that integrates longitudinal PROs with clinical features to improve survival prediction in HNC. Using CF-based imputation with iterative Bi-LSTM modeling, the framework enables forecasting of patient-specific symptom trajectories despite incomplete PROs and summarizes these trajectories into data-driven, cluster-level symptom burden representations for survival analysis.

Our results show that incorporating predicted long-term symptom burden improves prognostic discrimination beyond clinical variables alone, as reflected by gains in c-index and time-dependent AUC. Importantly, this approach enables symptom-informed risk assessment at clinically relevant decision points using baseline PROs and routinely collected clinical data. Together, these findings highlight the value of modeling longitudinal PRO dynamics and support the use of data-driven symptom burden representations as a practical bridge between trajectories and personalized survival modeling.