DITTO: A Visual Digital Twin for Interventions and Temporal Treatment Outcomes in Head and Neck Cancer

TL;DR

DITTO addresses the challenge of individualized risk assessment in head and neck cancer by constructing a sequential Deep Reinforcement Learning digital twin that predicts short- and long-term disease outcomes and toxicities. It combines transition models and Deep Survival Machines to model time-to-event outcomes, with optimal and imitation policy models built on a transformer-based encoder and a KNN symptom predictor to capture trajectory data, while dropout-based confidence intervals and integrated gradients provide clinical interpretability. The work reports quantitative performance on a 536-patient cohort and qualitative case studies, and distills design lessons for clinical visual XAI. Overall, the approach aims to assist clinicians in balancing tumor control with toxicity and mortality risks by delivering personalized, interpretable risk profiles and treatment recommendations.

Abstract

Digital twin models are of high interest to Head and Neck Cancer (HNC) oncologists, who have to navigate a series of complex treatment decisions that weigh the efficacy of tumor control against toxicity and mortality risks. Evaluating individual risk profiles necessitates a deeper understanding of the interplay between different factors such as patient health, spatial tumor location and spread, and risk of subsequent toxicities that can not be adequately captured through simple heuristics. To support clinicians in better understanding tradeoffs when deciding on treatment courses, we developed DITTO, a digital-twin and visual computing system that allows clinicians to analyze detailed risk profiles for each patient, and decide on a treatment plan. DITTO relies on a sequential Deep Reinforcement Learning digital twin (DT) to deliver personalized risk of both long-term and short-term disease outcome and toxicity risk for HNC patients. Based on a participatory collaborative design alongside oncologists, we also implement several visual explainability methods to promote clinical trust and encourage healthy skepticism when using our system. We evaluate the efficacy of DITTO through quantitative evaluation of performance and case studies with qualitative feedback. Finally, we discuss design lessons for developing clinical visual XAI applications for clinical end users.

Figures (11)

• Figure 1: Architecture for the transition and deep survival models (DSM). Patient state and previous state treatment decision use a standard DNN with input dropout to improve the models ability to deal with unknown data. The decision is concatenated to the penultimate layer in order to prevent the model from relying only on correlated features due to the use of dropout during training. DSM models predict a mixture of model parameters for each patient from a pre-trained set of user-defined number of mixtures.
• Figure 2: Architecture for policy model used to simulate a physician decision. Both the optimal and imitation models use a shared embedding with a custom position token at each stage, followed by a separate layer for each output, with additional fully connected layers unique to each model before the output. Model activations for the penultimate layers are used when calculating similar patients. Policy models use a modified version of a transformer encoder that saves the cohort at each time point into memory at training time.
• Figure 3: All Transition State Outcomes. (Right) Accuracy and AUC score for boolean outcomes such as toxicities. Models perform well in terms of accuracy and late toxicity (FT and Aspiration), but have mixed AUC results for dose-limiting toxicities due to the heavy imbalance in the data and low number of positive samples to learn from. (Center) Model performance for multi-class transition states (disease response and dose modification) using accuracy and micro, macro, and weighted AUC score for both unweighted and balanced loss weights. Models perforce best in terms of macro AUC score. Balanced models generally performed worse. (Right) F1 score and AUC score for temporal outcomes at 12, 24, 36, and 48 months after treatment. F1 scores tend to be very high while AUC scores stay around .6, likely due to issue with imbalanced data and incomplete censoring.
• Figure 4: Scatterplot used during model development using the 2 principal components of the policy model embeddings of the cohort and main patient. Outer and inner color encodes model predicted treatment and ground truth treatment, respectively. Hue is used to differentiate the current patient, the most similar patients, and the rest of the cohort.
• Figure 5: Diagrams used for spatial features in the visualization. (Left) Dose-limiting toxicities. (Center) Lymph node regional involvement. (Right) Primary tumor subsite. All regions not included in the diagram are considered "Not Otherwise Specified".