System and Method for Distributionally Robust Training of Physics-Constrained Cancer Digital Twins Using Dual-Mode Risk Set Construction

A computer-implemented method for training a distributionally robust cancer survival prediction model, comprising:

1. receiving, by a processor, a training dataset comprising multi-modal patient data and associated patient identifier strings;
2. parsing, by the processor, the patient identifier strings to extract a Tissue Source Site (TSS) code for each patient, wherein the TSS code indicates a data acquisition environment;
3. assigning each patient to one of a plurality of environment groups ($e_k$) based on the extracted TSS code;
4. executing, by a GPU, a forward pass of a neural network to generate risk scores and physics-constrained tumor dynamics parameters for the patients;
5. computing, for each environment group $e_k$, a partial log-likelihood loss ($L_k$) by applying a configured risk set construction protocol, wherein said protocol defines which patients in a batch constitute a risk set denominator for a given patient;
6. updating a set of environment weights $q_k$ using an exponentiated gradient ascent rule defined as $q_k \leftarrow q_k \cdot \exp(\eta \cdot L_k) / Z$, wherein $\eta$ is a step size and $Z$ is a normalization factor; and
7. updating model parameters of the neural network to minimize a weighted sum of the per-environment partial log-likelihood losses weighted by the updated environment weights $q_k$.

The method of claim 1, wherein the configured risk set construction protocol is a pooled cross-site protocol comprising: computing, from risk scores of patients from all environment groups in the batch, a global cumulative log-sum-exp term defining a global risk denominator for each event time; and computing the partial log-likelihood loss $L_k$ by summing log-likelihood terms strictly for event observations assigned to environment group $e_k$, such that the risk denominator for a patient assigned to environment group $e_k$ includes contributions from patients assigned to at least one other environment group, thereby enforcing cross-site ranking invariance.

The method of claim 1, wherein the configured risk set construction protocol is a strictly isolated protocol comprising:

1. allocating, in a GPU memory, a contiguous hazard buffer $B$ partitioned into $K$ segments, wherein each segment $B_k$ corresponds to an environment group $e_k$ and spans indices $[\text{start}_k,\, \text{end}_k)$;
2. constructing a segment descriptor array $S$ containing tuples of $(\text{start}_k,\, \text{end}_k,\, n\_\text{events}_k)$; and
3. executing a segmented logcumsumexp reduction kernel parameterized by the segment descriptor array $S$, wherein said kernel resets accumulation at segment boundaries defined by $S$, such that the risk denominator for a patient in environment group $e_k$ is computed exclusively from patients within the same environment group $e_k$.

The method of claim 3, wherein the segmented logcumsumexp reduction kernel structurally prevents cross-environment memory access during denominator accumulation, thereby ensuring per-environment statistical independence.

The method of claim 1, wherein the physics-constrained tumor dynamics parameters include a tumor growth rate ($\rho$) constrained to the range $[0, 0.3]$ per day via a sigmoid scaling function, and a drug sensitivity coefficient ($\beta$) constrained to the range $[0, 1]$.

The method of claim 1, further comprising merging any environment group containing fewer than a predetermined threshold of samples into a composite environment group prior to computing the partial log-likelihood loss.

A computer-implemented system for personalized oncology therapy selection, comprising: a memory storing a trained multi-modal neural network model; and a processor coupled to the memory and configured to:

1. receive multi-modal input data for a target patient, including omics data and optional whole slide imaging (WSI) data;
2. generate, using the trained multi-modal neural network model, a set of physics-constrained tumor dynamics parameters comprising a growth rate ($\rho$), a drug sensitivity ($\beta$), and an immune kill rate ($\omega$);
3. enforce biological constraints on the parameters such that $\rho \in [0, 0.3]$, $\beta \in [0, 1]$, and $\omega > 0$;
4. receive a definition of a plurality of candidate therapeutic regimens, wherein each regimen specifies a drug concentration function $D(t)$ over a time horizon;
5. for each candidate therapeutic regimen, simulate a tumor volume trajectory $V(t)$ by numerically integrating a differential equation parameterized by the generated physics-constrained parameters and the drug concentration function $D(t)$;
6. select a recommended therapeutic regimen from the plurality of candidate therapeutic regimens that has a minimum predicted tumor burden at a specified clinical time horizon;
7. generate a treatment recommendation data structure comprising at least: (i) a selected_regimen_id identifying the recommended therapeutic regimen; (ii) a predicted_trajectory_array of dimensions $[T \times S]$; (iii) an uncertainty_bounds array of dimensions $[T \times S \times 2]$ derived from Monte Carlo dropout forward passes; and (iv) physics_compliance_flags indicating adherence to the biological constraints; and
8. store the treatment recommendation data structure in a database record associated with a patient identifier for the target patient, or transmit the treatment recommendation data structure to a clinical decision support interface configured to display the predicted trajectory array.

The system of claim 7, wherein the trained multi-modal neural network model is trained using Distributionally Robust Optimization (DRO) over hospital-specific environment groups to ensure the generated parameters are robust to data acquisition artifacts.

The system of claim 7, wherein the processor is further configured to:

1. calculate an Information Sufficiency Score (ISS) for the target patient based on a distance between the target patient's latent representation and a training distribution centroid; and
2. route the multi-modal input data to a robust generalist checkpoint of the neural network model if the ISS exceeds a sufficiency threshold, or to a fallback checkpoint if the ISS is below the sufficiency threshold.

The system of claim 7, wherein numerically integrating the differential equation comprises using a fixed-step Euler solver or an adaptive Runge-Kutta solver.

The system of claim 7, wherein the differential equation is a Lotka-Volterra equation of the form:

wherein $V$ is tumor volume, $K_{\text{cap}}$ is carrying capacity, and $I(t)$ is an immune state function.

The system of claim 7, wherein the uncertainty_bounds are computed by executing a plurality of forward passes of the neural network with dropout enabled, and calculating the 2.5th and 97.5th percentiles of the resulting trajectory distribution at each time step.

A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to perform operations comprising:

1. extracting environment labels from patient barcodes in a multi-center cancer dataset, wherein the environment labels correspond to Tissue Source Sites (TSS);
2. assigning patients to environment groups based on the extracted environment labels;
3. executing a forward pass of a neural network to predict tumor dynamics parameters;
4. computing a per-environment partial log-likelihood loss ($L_k$) for each environment group using a pooled risk set construction, wherein said pooled risk set construction comprises: calculating a global cumulative log-sum-exp term across all environment groups in a batch; and masking event contributions such that $L_k$ aggregates loss only for patients in environment group $k$ while utilizing the global cumulative log-sum-exp term as a denominator, thereby enforcing cross-site ranking invariance;
5. updating a set of environment weights $q_k$ using an exponentiated gradient ascent rule: $q_k \leftarrow q_k \cdot \exp(\eta \cdot L_k) / Z$; and
6. updating model parameters of the neural network to minimize a weighted sum of the per-environment partial log-likelihood losses weighted by $q_k$.

The non-transitory computer-readable medium of claim 13, wherein the operations further comprise projecting the predicted tumor dynamics parameters into a bounded feasible set prior to computing the loss, ensuring that the predicted parameters remain physically valid regardless of the environment weights.

The non-transitory computer-readable medium of claim 13, wherein the neural network comprises a Variational Autoencoder (VAE) for genomic encoding and a separate hypernetwork for parameter prediction, and wherein the VAE is frozen during the updating of the neural network.

The non-transitory computer-readable medium of claim 13, wherein the pooled risk set construction enables the neural network to learn biological features that rank patients consistently across different Tissue Source Sites.

The non-transitory computer-readable medium of claim 13, wherein the patient barcodes are formatted according to The Cancer Genome Atlas (TCGA) standards, and the TSS is extracted from a fixed position within the barcode string.