Generative Deep Patient Models

• Generative Deep Patient (GDP) systems are AI models that fuse structured EHR data with unstructured clinical text to generate patient-centric records.
• They leverage diverse architectures—from LLaMA-based decoders to recurrent networks—to support note generation, risk prediction, and trajectory simulation.
• Despite varied implementations, GDP systems share a focus on dynamic, longitudinal patient representation and iterative clinical documentation.

Generative Deep Patient (GDP) is an ambiguous term in recent machine-learning and clinical informatics literature. In the most direct EHR-specific usage, it denotes a multimodal foundation model that natively encodes structured EHR time-series and fuses them with unstructured clinical text into a LLaMA-based decoder for both clinical prediction and narrative generation (Sivarajkumar et al., 22 Aug 2025). More broadly, the label has been used descriptively for systems that generate longitudinal, patient-centric clinical notes from encounter data (Biswas et al., 2024), for generative latent-state models of longitudinal patient trajectories (Brouwer et al., 2018), and for EHR generators that synthesize patient cohorts for augmentation or data sharing (Choi et al., 2017, Che et al., 2017). The term is not stable across papers: other works expand GDP as “General Demographic Pre-trained” (Chen et al., 9 Sep 2025) or “Generative Distribution Prediction” (Tian et al., 10 Feb 2025), so its meaning depends on context.

1. Terminology and scope

The acronym GDP is not uniquely associated with a single architecture or task family. In clinical AI, three distinct usages appear in the supplied literature.

Usage of “GDP”	Definition	Paper
Generative Deep Patient	Multimodal foundation model for structured and unstructured EHRs	(Sivarajkumar et al., 22 Aug 2025)
General Demographic Pre-trained	Foundation-style encoder for age and gender	(Chen et al., 9 Sep 2025)
Generative Distribution Prediction	Unified multimodal prediction framework based on estimated conditional distributions	(Tian et al., 10 Feb 2025)

The 2025 demographic paper is explicit that its GDP is a General Demographic Pre-trained model and that calling it “Generative Deep Patient” would be misleading; it is described as “primarily discriminative / representation-based” and “not a generative model” in the sense of VAEs, GANs, autoregressive EHR simulators, or generative Deep Patient extensions (Chen et al., 9 Sep 2025). Likewise, the multimodal learning paper uses GDP for Generative Distribution Prediction, a modality-agnostic framework for tabular, text, and image prediction rather than an EHR-specific patient model (Tian et al., 10 Feb 2025).

Within EHR research proper, “Generative Deep Patient” is therefore best understood as a family resemblance term rather than a single canonical method. It usually implies a patient-level representation that is both deep and generative: the model either generates structured notes, future clinical events, synthetic records, or grounded clinical narratives while maintaining a patient-centric state across time (Biswas et al., 2024, Brouwer et al., 2018, Sivarajkumar et al., 22 Aug 2025).

2. Patient representation paradigms

A defining property of GDP-style systems is the representation of patient state as an object that can be updated, decoded, or rolled forward. The supplied literature presents several such representations.

In clinical documentation pipelines, the patient is represented encounter by encounter through structured SOAP or BIRP notes. The 2024 documentation paper describes a pipeline in which diarized transcripts are converted into SOAP/BIRP notes, including structured JSON fields such as subjective.chiefComplaint, objective.vitalSigns, progressAndResponse, homework, and planForFutureSession. The same paper describes iterative note improvement across visits, producing Structured Clinical Note V1, V2, and later versions as new transcripts and documents arrive (Biswas et al., 2024). This suggests a GDP-like patient state implemented as a versioned, patient-centric documentation object rather than as a latent vector.

In longitudinal generative trajectory modelling, the patient is represented by a latent temporal process. “Deep Ensemble Tensor Factorization for Longitudinal Patient Trajectories Classification” models an order-3 tensor $\mathcal{Y} \in \mathbb{R}^{N \times M \times T}$ and assigns each patient a hidden health state $\mathbf{h}^i[t] \in \mathbb{R}^{D}$ that evolves over time. Observed measurements are generated as

$$\mathcal{Y}_{i,j,t} = g^j\left(\mathbf{h}^i[t]\right) + \epsilon,$$

and the label is generated from the final latent state through $p(z_i = 1) = w(\mathbf{h}^i[T])$ (Brouwer et al., 2018). Here the “deep patient” representation is the trajectory $\{\mathbf{h}^i[t]\}$, with $\mathbf{h}^i[T]$ serving as a compact summary for downstream classification.

In multimodal foundation modelling, the patient representation is fused across structured time series, demographics, and text. The 2025 GDP foundation model encodes a structured EHR timeline with a CNN-Transformer encoder, adds a 256-dimensional demographic embedding to each time step, projects a BioClinicalBERT note embedding from 768 to 256 dimensions, and uses the resulting fused sequence as cross-attention memory for a 3.2B-parameter LLaMA-like decoder (Sivarajkumar et al., 22 Aug 2025). For clinical prediction, the hidden state at a special <BOS> token is taken as the patient representation and fed to task-specific heads for heart failure, type 2 diabetes, and 30-day readmission (Sivarajkumar et al., 22 Aug 2025).

A related representation appears in Foresight, where each patient is converted into a sequence of SNOMED CT concept tokens extracted from free text, augmented with demographics, age-change tokens, and <SEP> delimiters between daily buckets. The model is trained autoregressively over sequences

$$p_\theta(x_{1:T}) = \prod_{t=1}^{T} p_\theta(x_t \mid x_{<t}),$$

so the transformer history serves as a representation of the patient timeline at each generation step (Kraljevic et al., 2022). This suggests a GDP-like formulation in which the patient state is implicit in the causal transformer context rather than explicitly parameterized.

3. Architectural patterns and training objectives

GDP-style systems span several architectural regimes, but most combine longitudinal encoding with constrained generation.

The documentation-oriented pipeline in “Intelligent Clinical Documentation” is explicitly multi-stage. Data collection begins from audio or video of patient–clinician interactions; ASR produces a raw transcript; utterance classification assigns each utterance to “patient” or “clinician”; prompting then converts the diarized transcript into SOAP or BIRP notes; and structured output can be enforced with function calling and a JSON Schema (Biswas et al., 2024). The diarization stage is framed as binary sequence labeling with softmax and cross-entropy, while note generation follows standard conditional generation,

$$P(y \mid x) = \prod_{t=1}^{T} P(y_t \mid y_{<t}, x).$$

The paper does not describe fine-tuning; instead it uses prompt engineering, zero-shot or one-shot prompting, and schema-constrained generation with models including GPT‑3.5, GPT‑4 TURBO, Claude v3, Mixtral‑8x7B Instruct, and Llama‑3 70B Instruct (Biswas et al., 2024).

The multimodal GDP foundation model introduced in 2025 has a more integrated encoder–decoder design. Its structured encoder takes a [T × 128] patient timeline, applies three temporal convolutions, then a 4-layer Transformer encoder with hidden dimension 256, 8 attention heads, FFN inner dimension 512, and dropout 0.1. Its decoder is a 24-layer LLaMA-like model with masked self-attention, cross-attention to fused EHR embeddings, and a 256 → 1024 → 256 FFN (Sivarajkumar et al., 22 Aug 2025). Pretraining combines three losses:

$$\mathcal{L}_{\text{pretrain}} = \mathcal{L}_{\text{LM}} + \lambda_{\text{MFP}} \mathcal{L}_{\text{MFP}} + \lambda_{\text{NTP}} \mathcal{L}_{\text{NTP}},$$

where $\mathcal{L}_{\text{LM}}$ is conditional language modeling on Brief Hospital Course text, $\mathbf{h}^i[t] \in \mathbb{R}^{D}$0 is masked feature prediction, and $\mathbf{h}^i[t] \in \mathbb{R}^{D}$1 is next time-step prediction (Sivarajkumar et al., 22 Aug 2025).

The latent-trajectory branch of GDP uses recurrent generative dynamics. In the 2018 tensor factorization model, the GRU receives an input vector concatenating observed values and the observation mask, with missing entries imputed by the model’s own predictions from the previous time step. Training minimizes a weighted combination of reconstruction, cross-entropy classification, and $\mathbf{h}^i[t] \in \mathbb{R}^{D}$2 regularization:

$\mathbf{h}^i[t] \in \mathbb{R}^{D}$3

An ensemble of 200 models is trained, with the best 20 selected by validation performance to mimic Bayesian posterior sampling (Brouwer et al., 2018).

GAN-based GDP precursors adopt another pattern. medGAN learns a continuous latent patient embedding with an autoencoder and then trains a GAN to generate high-dimensional discrete binary or count records through the decoder, with minibatch averaging introduced to reduce mode collapse (Choi et al., 2017). ehrGAN instead learns a transition distribution $\mathbf{h}^i[t] \in \mathbb{R}^{D}$4 over EHR sequences and uses a variational contrastive divergence-style generator to create synthetic neighbors of real patients for semi-supervised augmentation (Che et al., 2017).

4. Documentation, narrative generation, and longitudinal note synthesis

One major interpretation of GDP treats generation itself as a documentation interface to the patient record. The 2024 note-generation paper focuses on SOAP and BIRP draft creation from transcribed encounters and describes basic prompting, advanced prompting, and structured prompting with JSON schema enforcement (Biswas et al., 2024). Its case study uses simulated therapy sessions from YouTube, Whisper-style ASR, diarization via GPT‑3.5 when pyannotate diarization was insufficiently accurate, and machine-readable JSON outputs designed for downstream EHR integration (Biswas et al., 2024).

The same paper emphasizes iterative note refinement across visits. New transcripts and documents are combined with an existing note, and the LLM is prompted to generate an updated note that reflects both the existing state and newly observed information (Biswas et al., 2024). This supports a longitudinal view in which the patient record is not regenerated de novo at each visit but revised as a versioned state. A plausible implication is that GDP, in this usage, is less a standalone predictive model than a continuously updated clinical representation anchored in documentation.

The multimodal GDP foundation model generalizes this narrative capacity from encounter transcripts to inpatient EHRs. During pretraining it generates the Brief Hospital Course section of discharge summaries conditioned on structured timelines and note embeddings, optimizing

$\mathbf{h}^i[t] \in \mathbb{R}^{D}$5

For narrative generation it reports ROUGE‑L = 0.135 and BERTScore‑F1 = 0.545, and in blinded human evaluation GDP‑Instruct scored highest on faithfulness, fluency, and overall clinical utility (Sivarajkumar et al., 22 Aug 2025).

The documentation paper’s internal evaluation is narrower but technically informative. On a dataset of 20 SOAP and 20 BIRP notes, GPT‑4 achieved ROUGE‑1 F1 ~ 0.90–0.95 across SOAP-note sample complexity, while Claude v3 and Llama‑3 70B scored ~ 0.70–0.80, and Mixtral‑8x7B scored ~ 0.65–0.75; similar patterns held for BIRP notes (Biswas et al., 2024). The paper also states that notes were “graded for quality by humans,” although it does not give a detailed human-rating scale (Biswas et al., 2024).

Taken together, these systems move GDP beyond latent representation learning. They make the patient model legible as both structured JSON and clinician-facing prose, which is central to EHR deployment because the model’s state can enter workflows before, during, or after visits and remain subject to clinician review and sign-off (Biswas et al., 2024).

5. Prediction, simulation, and synthetic patient generation

GDP-style models are also used for risk prediction, event forecasting, synthetic cohort generation, and semi-supervised augmentation.

In the multimodal GDP foundation model, clinical prediction on MIMIC‑IV yielded heart failure AUROC = 0.923, type 2 diabetes AUROC = 0.817, and 30-day readmission AUROC = 0.627 (Sivarajkumar et al., 22 Aug 2025). Supplementary ablations reported that removing next time-step prediction reduced heart failure AUROC to 0.893, removing masked feature prediction reduced it to 0.915, and removing both reduced it to 0.870, indicating that temporal self-supervision materially contributes to discrimination (Sivarajkumar et al., 22 Aug 2025).

The 2018 deep ensemble tensor factorization model addresses scarcely observed longitudinal ICU trajectories. On in-hospital mortality prediction from 96 longitudinal measurement types during the first 48 hours after admission, the single best generative GRU model achieved AUC ≈ 0.842 and the ensemble achieved AUC ≈ 0.855, outperforming both SAPS‑II and a GRU-imputed baseline (Brouwer et al., 2018). Because the model reconstructs trajectories as well as predicts mortality, its latent state can be interpreted as a compact patient representation suitable for other downstream tasks.

Foresight shifts the target from binary risk to open-ended forecasting. It uses a GPT‑2–style transformer over SNOMED concept timelines and reports precision@10 for forecasting the next disorder of 0.68 at King’s College Hospital, 0.76 at South London and Maudsley, and 0.88 on MIMIC‑III; for forecasting the next biomedical concept, precision@10 was 0.80, 0.81, and 0.91, respectively (Kraljevic et al., 2022). The model can also forecast follow-on biomedical concepts “for as many steps as required” and was validated on 34 synthetic patient timelines by five clinicians, with relevancy of 97% for the top forecasted candidate disorder (Kraljevic et al., 2022).

Synthetic EHR generation forms another GDP strand. medGAN generates high-dimensional discrete patient records with a hybrid autoencoder-GAN and reports that synthetic records achieve comparable performance to real data on distribution statistics, predictive modeling tasks, and a medical expert review, while showing a limited privacy risk in both identity and attribute disclosure (Choi et al., 2017). ehrGAN, by contrast, uses generated trajectories as labeled augmentations. In limited-label settings it improved AUROC from 0.8784 to 0.9075 on HF50, from 0.9156 to 0.9354 on Dia50, from 0.8953 to 0.9246 on HF67, and from 0.9386 to 0.9563 on Dia67 (Che et al., 2017).

These results imply that GDP is not restricted to one output type. Depending on the formulation, it can produce a risk score, a synthetic record, a future disorder sequence, a discharge narrative, or a versioned SOAP/BIRP note. What remains common is the generative use of patient history rather than a purely static discriminative mapping.

6. Limitations, governance, and future directions

The literature consistently treats GDP-style systems as high-utility but safety-critical. In documentation generation, privacy and data protection are primary concerns: prompt templates explicitly instruct models to “Maintain patient confidentiality by avoiding the use of any personally identifiable information,” and the paper emphasizes risks of unauthorized access, breaches, bias, hallucinations, and liability, together with the need for clinician review, validation, and sign-off (Biswas et al., 2024). The same work treats human oversight as essential and frames AI as assistive rather than autonomous (Biswas et al., 2024).

The multimodal GDP foundation model lists substantial technical limits. Its context is restricted to ~100 time steps and one main note; it is trained only on MIMIC‑IV; it evaluates a limited task set; it does not yet include imaging, genomics, or waveforms; and it is “not inherently interpretable” despite the use of attention (Sivarajkumar et al., 22 Aug 2025). The paper also highlights bias and fairness risks, hallucinations, privacy constraints, and over-reliance on automation, and suggests longer-context retrieval, more modalities, RLHF and safety tuning, adapter-based domain adaptation, and prospective clinical evaluation as future directions (Sivarajkumar et al., 22 Aug 2025).

Foresight adopts an even more explicit cautionary stance. It warns against clinical decision-support deployment in its current form because it learns historical common practice, prioritizes probability over urgency or impact, and can hallucinate like other GPT-like systems (Kraljevic et al., 2022). It also inherits biases from EHR sampling, documentation order, and NER/linking errors, while truncating timelines to 256 concepts and removing concepts occurring fewer than 100 times globally (Kraljevic et al., 2022).

A conceptual controversy concerns the word “generative” itself. In the multimodal foundation model (Sivarajkumar et al., 22 Aug 2025), the term refers to clinical narrative generation grounded in fused EHR representations. In the documentation pipeline (Biswas et al., 2024), it refers to LLM-based generation of notes and iterative updates from transcripts. In the trajectory and GAN literature (Brouwer et al., 2018, Choi et al., 2017, Che et al., 2017), it refers to explicit generation of measurements, latent states, or synthetic patient records. By contrast, the demographic GDP model explicitly rejects a generative interpretation, underscoring that GDP is not a settled technical term across subfields (Chen et al., 9 Sep 2025).

The most stable synthesis is therefore architectural rather than lexical. A Generative Deep Patient system, in the EHR sense, is a patient-level model that maintains a deep representation across time and uses that representation to generate clinically meaningful artifacts—notes, trajectories, forecasts, or synthetic records—under explicit constraints of grounding, review, and governance (Biswas et al., 2024, Brouwer et al., 2018, Sivarajkumar et al., 22 Aug 2025).