Generative Diagnosis Paradigm

• The generative diagnosis paradigm is a diverse set of methods where generation is central to creating diagnostic evidence, such as augmenting data, reconstructing missing modalities, or inferring latent states.
• It enhances diagnostic accuracy and fairness by synthesizing underrepresented samples, generating diagnostic hypotheses, and guiding inference with structured generative processes.
• Applications include dermatology, neuroimaging, lab tests, and industrial fault diagnosis, with reported improvements in F1, AUC, and robustness across various clinical and non-clinical settings.

Searching arXiv for papers relevant to the generative diagnosis paradigm and the cited examples. The generative diagnosis paradigm is a family of diagnostic formulations in which generation is not peripheral to inference, but a central mechanism for producing diagnostic evidence, latent states, missing modalities, balanced training distributions, or structured hypothesis sets. Across the literature, this paradigm departs from purely discriminative mappings such as $Y=f(X)$ or $P(Y\mid X)$ and instead uses generative components to support diagnosis in several distinct ways: by augmenting underrepresented diagnostic strata, by reconstructing diagnostically meaningful but difficult-to-measure modalities, by inferring latent patient states from incomplete trajectories, by autoregressively generating ranked differentials, by making modality synthesis diagnosis-aware, or by modeling joint medical variables and performing test-time inference under partial observation (Munia et al., 21 Mar 2025, Na et al., 2023, Zhang et al., 2017, Alam et al., 2023, Shin et al., 2020, Zhang et al., 11 May 2026). A useful Editor’s term for these diverse instantiations is “generation-for-diagnosis”: diagnosis may still terminate in a conventional classifier or decision rule, but generation reshapes the evidence, representation, or search space from which the diagnosis is made.

1. Conceptual scope and defining forms

The paradigm is best understood as plural rather than singular. The papers in this area do not all “diagnose by generating” in the same way. Instead, they instantiate several recurring formulations.

One formulation is generative augmentation for diagnosis, where a model synthesizes new training data to improve a downstream diagnostic classifier. In dermatology, DermDiff is explicitly described as a generative, data-centric approach to diagnosis rather than a model that diagnoses by generation itself. It uses text-conditioned latent diffusion to generate balanced dermoscopic images across skin-tone and lesion-status strata, with the downstream endpoint remaining benign-vs-malignant classification (Munia et al., 21 Mar 2025). A related but distinct case appears in schizophrenia diagnosis from EEG, where class-specific VAE or WGAN-GP generators synthesize spectrograms to reduce overfitting and improve a CNN classifier; the paper itself characterizes this as diagnosis strengthened by generation rather than a fully generative diagnostic framework (Saadatinia et al., 2023). SkinGenBench generalizes this logic to melanoma diagnosis and explicitly treats image synthesis as useful only insofar as it improves downstream clinical classification, especially melanoma F1 and ROC-AUC under malignant-class scarcity (Pritam et al., 19 Dec 2025).

A second formulation is evidence-generating diagnosis or virtual sensing, where the model generates a diagnostically meaningful but hard-to-measure modality from an easier one. In critical heat flux diagnosis, a conditional GAN reconstructs infrared thermal fields from total reflection images of boiling interface behavior, so that CHF is inferred from the generated temperature distribution and irreversible dry-patch evolution rather than from a direct class label (Na et al., 2023). In Alzheimer’s disease imaging, GANDALF synthesizes PET-like information from MRI, but unlike earlier synthesize-then-classify pipelines, it injects diagnosis loss directly into GAN training so that synthetic PET is optimized for diagnostic utility as well as realism (Shin et al., 2020).

A third formulation is latent-state generative diagnosis, where the model explains partially observed clinical measurements through hidden stochastic states and diagnoses from those states. The VRNN+NN model for laboratory-test diagnosis is paradigmatic here: the generative recurrent model handles longitudinal missingness and uncertainty, while the discriminative head maps the inferred hidden-state sequence to ICD-9 diagnoses (Zhang et al., 2017). The diagnosis is therefore mediated by latent physiological trajectories rather than by a direct mapping from sparse zero-filled inputs.

A fourth formulation is generative differential diagnosis, in which the model directly generates ranked disease hypotheses. DDxT recasts differential diagnosis as conditional sequence generation over pathology tokens, replacing reinforcement-learning evidence-acquisition policies with a Transformer that autoregressively emits an ordered differential and then predicts the final pathology using pooled encoder and decoder representations (Alam et al., 2023).

A fifth formulation is joint generative reformulation, where the task is redefined from $X\to Y$ prediction to joint modeling of $P(X,Y)$ followed by test-time conditioning or optimization. GenMed is the clearest expression of this idea in medical imaging: a diffusion model is trained over paired medical variables and, at test time, reverse diffusion is guided so that generated samples match whatever subset of variables is observed, enabling segmentation, cross-modality transfer, degraded-input inference, and shape completion without retraining (Zhang et al., 11 May 2026).

A plausible implication is that the generative diagnosis paradigm is best defined functionally: generation is central whenever it changes how diagnostic evidence is represented, completed, balanced, or searched, even if the final decision remains discriminative.

2. Generative augmentation as data-centric diagnosis

In data-centric variants, the generator intervenes upstream on the training distribution. The clearest medical example is DermDiff, which addresses a concrete fairness failure mode in dermatology: lesion classifiers are often trained on highly imbalanced datasets dominated by lighter skin tones and underperform on darker skin tones. The paper explicitly frames this as both a dataset imbalance problem and a racial/skin-tone bias problem, and positions generative modeling as a way to expand the training distribution, especially for rare combinations of disease status and skin tone (Munia et al., 21 Mar 2025).

Technically, DermDiff is a text-conditioned latent diffusion framework built by fine-tuning Stable Diffusion v1.4 with 1.45B parameters originally trained on LAION-400M. The conditioning variables are deliberately limited: prompts are derived from metadata and encode only skin tone and disease type/status. The paper gives the denoising relation

$$\tilde{\epsilon} = UNet(VAE(x)\oplus\epsilon, t, CLIP(y)),$$

with training objective

$$\mathcal{L}_{mse} = \frac{1}{d} \sum_{i=0}^{d-1} (\tilde{\epsilon}_{i} - \epsilon_{i})^2.$$

No custom reverse-update formula, $\beta_t/\alpha_t$ schedule, classifier-free guidance weight, or custom variational objective is specified in the paper text (Munia et al., 21 Mar 2025).

A distinctive element is that subgroup conditioning depends on a separate skin-tone detector. Fitzpatrick types are collapsed into three groups—A: Fitzpatrick I–II, B: III–IV, C: V–VI—and detected with ResNeXt-101 trained with focal loss

$$\mathcal{L}_{focal} = -\alpha (1 - \hat{p})^\gamma \log(\hat{p}),$$

using $\alpha=[0.3,0.4,0.9]$ and $\gamma=2$ for skin-tone detection (Munia et al., 21 Mar 2025). This pipeline matters because many dermatology datasets lack reliable demographic labels; subgroup analysis on several external datasets therefore depends on predicted skin-tone labels rather than human annotations.

The synthetic-data intervention is explicit. DermDiff generates 60,000 synthetic dermoscopic images: 30,000 benign and 30,000 malignant, with 10,000 images for each of the three skin-tone groups within each disease class. These images are merged with real data to train downstream lesion classifiers based again on ResNeXt-101. On DDI, the real+synthetic setup improves F1 and AUC across all skin tones relative to Fitzpatrick-only training, including group C where AUC rises from 0.51 to 0.58 and F1 from 0.49 to 0.52 (Munia et al., 21 Mar 2025). The paper also claims improvements in accuracy parity, equal opportunity, and equalized odds for darker skin tones, though numeric values for those parity metrics are not reported.

Other augmentation-oriented diagnosis papers reinforce the same logic under different modalities. In schizophrenia diagnosis from EEG spectrograms, VAE-generated class-specific spectrograms improved test accuracy from 96.0% to 99.0%, F1 from 0.960 to 0.990, sensitivity from 0.958 to 0.992, and specificity from 0.962 to 0.986 on the 16-channel dataset (Saadatinia et al., 2023). The paper emphasizes that the VAE synthetic distribution is closer to the original data than the WGAN-GP distribution, as reflected by TSTR/TRTS scores of 94.8%/98.1% for the VAE versus 80.5%/80.5% for WGAN (Saadatinia et al., 2023). In SkinGenBench, synthetic melanoma augmentation yields 8–15% absolute gains in melanoma F1-score, with ViT-B/16 achieving F1 $P(Y\mid X)$0 and ROC-AUC $P(Y\mid X)$1, and the paper concludes that generative architecture choice matters more than preprocessing complexity (Pritam et al., 19 Dec 2025).

These results collectively suggest that in data-limited or subgroup-imbalanced settings, the generative diagnosis paradigm often functions as a distribution repair mechanism. Its efficacy depends less on abstract perceptual realism than on class faithfulness, subgroup anchoring, and the extent to which synthetic samples alter the training distribution in diagnostically useful ways.

3. Diagnosis by modality translation and virtual sensing

A more direct generative formulation treats the diagnostic object itself as a generated hidden field or missing modality. The CHF work is exemplary because it reframes diagnosis as conditional generation of latent thermal evidence. Instead of predicting a binary CHF label, the model learns a map

$P(Y\mid X)$2

where $P(Y\mid X)$3 is a total reflection image and $P(Y\mid X)$4 the corresponding infrared thermometry image (Na et al., 2023). The diagnostic rationale is physical: phase-interface and dry-patch behavior visible in total reflection images is coupled to local surface heating, so reconstructing the thermal field yields evidence about CHF onset.

The architecture is a Pix2Pix-style conditional GAN with a modified U-Net generator and modified PatchGAN discriminator. The paper writes the conditional adversarial objective and its $P(Y\mid X)$5-augmented minimax form, and uses paired TR–IR images acquired in synchronized boiling experiments (Na et al., 2023). The latent bottleneck is reported as

$P(Y\mid X)$6

and the discriminator explicitly relies on patchwise local realism consistent with a Markov random field assumption. Reconstruction quality is measured by PSNR and SSIM, with Run 1 achieving PSNR = 23.73, SSIM = 0.7476, and Run 2 achieving PSNR = 22.49, SSIM = 0.7277 (Na et al., 2023). More importantly, generated thermal maps reproduce average temperature trends well, though maximum temperature is underestimated by 3.5% in Run 1 and 3.3% in Run 2. CHF evidence is then extracted by applying a threshold of $P(Y\mid X)$7 to identify irreversible dry patches (Na et al., 2023).

GANDALF pursues a related strategy in neuroimaging but makes synthesis explicitly diagnosis-aware. Earlier MRI-to-PET approaches generated PET and trained a separate classifier afterward. GANDALF instead incorporates Alzheimer’s classification into the GAN objective, using a pix2pix-style conditional GAN where the discriminator also predicts disease labels from MRI-PET pairs (Shin et al., 2020). The base objective is

$P(Y\mid X)$8

with

$P(Y\mid X)$9

and

$X\to Y$0

The paper then augments both generator and discriminator objectives with diagnosis losses and introduces discriminator-adaptive loss schedules to shift emphasis from realism to classification as training progresses (Shin et al., 2020).

The empirical pattern is task-dependent. GANDALF does not improve over the best MRI-only CNN on binary AD/CN, but it substantially improves harder multiclass settings: in the three-class AD/MCI/CN task, accuracy rises to 78.7 and $X\to Y$1 to 0.69, outperforming the synthesize-then-classify baseline at 71.3 and 0.63; in the four-class AD/LMCI/EMCI/CN setting, accuracy improves from 33.0 to 37.0 and $X\to Y$2 from 0.34 to 0.40 (Shin et al., 2020). This matters conceptually: the generated modality is diagnostically valuable not because it is merely realistic, but because it is optimized to amplify disease-relevant information.

These papers instantiate a strong version of generative diagnosis: the model generates the evidence field from which diagnosis is made. The advantage is interpretability at the level of generated thermal maps or PET-like volumes rather than opaque labels. The limitation is that generated evidence can still be quantitatively imperfect or clinically unvalidated.

4. Latent-state and sequence-generative diagnosis

Another branch of the paradigm uses generation to infer latent states or diagnosis sequences rather than images. In longitudinal laboratory diagnosis, the VRNN+NN model treats each patient record as a multivariate time series

$X\to Y$3

with substantial missingness. The key move is to learn diagnosis jointly with a generative temporal model of the lab sequence (Zhang et al., 2017). The VRNN defines

$X\to Y$4

updates hidden state via

$X\to Y$5

and predicts diagnosis from

$X\to Y$6

Its joint objective is

$X\to Y$7

with $X\to Y$8 (Zhang et al., 2017).

The clinical significance is that diagnosis is mediated by inferred latent physiological state rather than raw zero-filled observations. On MIMIC-derived data with 46,252 patients and the 50 most frequent diagnoses, VRNN+NN achieves Micro-F1 $X\to Y$9, Macro-F1 $P(X,Y)$0, and Micro-AUC $P(X,Y)$1, significantly outperforming the purely discriminative RNN+NN baseline at $P(X,Y)$2 across all metrics (Zhang et al., 2017). The model also yields better imputation, with MSE $P(X,Y)$3 compared with heuristic baselines such as last+next at $P(X,Y)$4. This supports the paper’s central claim that generative learning provides robustness to missing and noisy measurements while discriminative learning aligns the latent representation to diagnosis.

DDxT represents a more overtly generative sequence-based view. Given tokenized patient evidence, it autoregressively generates a ranked differential diagnosis using a Transformer encoder-decoder and then predicts the final pathology with a separate classifier (Alam et al., 2023). Although the paper does not print the factorization formula, the described mechanism corresponds to

$P(X,Y)$5

The architecture uses 6 encoder and 6 decoder Transformer blocks, hidden size 128, 4 attention heads, and vocabularies of size 436 for the encoder and 54 for the decoder (Alam et al., 2023). On DDXPlus, DDxT achieves GTPA@1 = 99.98, DDP = 94.84, DDR = 94.65, and DDF1 = 0.9472, outperforming RL-based baselines AARLC and BASD by a large margin (Alam et al., 2023).

These sequence- and latent-state formulations broaden the paradigm beyond image synthesis. Generation may produce latent patient trajectories or differential-diagnosis sequences; the common element is that diagnosis emerges from a model of how clinically meaningful states or hypothesis sequences are generated, not just from a direct discriminative mapping.

5. Joint modeling, test-time conditioning, and adaptive generative inference

A more ambitious reformulation models the joint distribution of medically meaningful variables and performs diagnosis as test-time inference under partial observation. GenMed explicitly contrasts this with classical discriminative learning $P(X,Y)$6 and instead trains diffusion models over

$P(X,Y)$7

where $P(X,Y)$8 and $P(X,Y)$9 are paired task variables such as image and mask (Zhang et al., 11 May 2026). The forward process follows standard DDPM notation,

$$\tilde{\epsilon} = UNet(VAE(x)\oplus\epsilon, t, CLIP(y)),$$0

and the reverse process is abstracted as a denoiser over paired variables. At inference, GenMed guides reverse diffusion by defining a consistency loss on the observed component and updating the current latent state by gradient descent before each denoising step: $$\tilde{\epsilon} = UNet(VAE(x)\oplus\epsilon, t, CLIP(y)),$$1

$$\tilde{\epsilon} = UNet(VAE(x)\oplus\epsilon, t, CLIP(y)),$$2

This enables arbitrary and previously unseen observation patterns without architectural changes or retraining (Zhang et al., 11 May 2026).

The empirical support is unusually broad. On MM-WHS and TotalSegmentator cardiac segmentation, GenMed-Full consistently outperforms nnU-Net, SwinUNETR, atlas-based methods, and an input-conditioned diffusion baseline. Under zero-shot CT-to-MRI transfer, it reaches average 59.91 Dice and 30.52 NSD versus 35.26/19.42 for nnU-Net; with only 2 CT training samples, it still achieves 77.67 Dice and 55.95 NSD (Zhang et al., 11 May 2026). Under severe degraded-input settings, GenMed-Full maintains strong performance where naive guidance collapses. This is diagnosis as posterior completion over a learned medical world model rather than prediction from a rigid input contract.

A different but conceptually related development is test-time instance-specific parameter composition, introduced in Composer for generative models (Tran et al., 29 Mar 2026). Composer is not a diagnostic paper, but it proposes a new paradigm of adaptive generation via one-shot low-rank weight updates: $$\tilde{\epsilon} = UNet(VAE(x)\oplus\epsilon, t, CLIP(y)),$$3 where $$\tilde{\epsilon} = UNet(VAE(x)\oplus\epsilon, t, CLIP(y)),$$4 and $$\tilde{\epsilon} = UNet(VAE(x)\oplus\epsilon, t, CLIP(y)),$$5 are generated per input condition. Adaptation occurs once before multi-step generation and improves class-conditional, text-to-image, and quantized generation with minimal overhead (Tran et al., 29 Mar 2026). The paper itself does not address diagnosis, but a plausible implication is that future generative diagnosis systems could use case-specific internal reparameterization rather than a single static generative prior for all patients. This suggests an eventual convergence between adaptive generative modeling and case-specific diagnostic world models.

Taken together, these works imply that the most general form of the generative diagnosis paradigm is not merely “generate extra data,” but “learn a reusable joint or adaptive generative model and perform inference by constraining or specializing it to the observed case.”

6. Reliability, interpretability, and limitations

A recurrent theme across the literature is that generative diagnosis introduces new forms of evidence and flexibility, but also new reliability problems. One major issue is that generated outputs can appear plausible while being diagnostically misleading. The CHF paper explicitly reports underestimation of maximum temperature and does not compare against simpler CNN regressors or direct classifiers (Na et al., 2023). GANDALF does not report image-fidelity metrics such as SSIM, PSNR, MAE, or FID, so its gains are tied to diagnosis-aware representation learning rather than demonstrated PET realism (Shin et al., 2020). DermDiff does not include expert dermatologist validation of synthetic lesion realism and stratifies several test sets by predicted, not human-annotated, skin-tone labels (Munia et al., 21 Mar 2025). SkinGenBench notes that advanced artifact removal may suppress clinically relevant textures despite modest metric gains (Pritam et al., 19 Dec 2025).

Another issue is double dipping in anomaly-based generative diagnosis. The Alzheimer’s selective-inference paper directly tackles this problem by pairing a healthy-only CVAE anomaly detector with post-selection hypothesis testing (He et al., 2024). The CVAE reconstructs a “healthy” alternative of optical-flow representations of longitudinal MRI change, and reconstruction error defines an anomaly mask. The paper then tests whether signal inside the selected anomaly region differs from the rest of the image, using

$$\tilde{\epsilon} = UNet(VAE(x)\oplus\epsilon, t, CLIP(y)),$$6

Naive p-values are shown to be invalid because the anomaly region was selected from the same data. Selective inference instead conditions on the selection event $$\tilde{\epsilon} = UNet(VAE(x)\oplus\epsilon, t, CLIP(y)),$$7, yielding calibrated p-values (He et al., 2024). Empirically, naive testing yields false discovery rate 0.74 on held-out normals, whereas selective inference controls FDR around the target level, with 0.06 at $$\tilde{\epsilon} = UNet(VAE(x)\oplus\epsilon, t, CLIP(y)),$$8 and power 0.33 versus 0.06 for Bonferroni (He et al., 2024). This is a notable advance because it transforms anomaly maps from heuristic outputs into statistically testable diagnostic objects.

Interpretability also varies by formulation. Evidence-generating systems such as CHF cGANs and MRI-to-PET synthesis produce inspectable fields or images (Na et al., 2023, Shin et al., 2020). VRNN-based lab diagnosis yields latent states that improve imputation but are not deeply mechanistically interpreted (Zhang et al., 2017). DDxT produces a ranked differential, which is more clinically legible than a single class prediction (Alam et al., 2023). GenMed returns anatomically coherent completions under arbitrary observation patterns, which is a strong form of structural interpretability (Zhang et al., 11 May 2026). By contrast, data-augmentation approaches improve fairness or F1 without necessarily offering case-level explanations (Munia et al., 21 Mar 2025, Pritam et al., 19 Dec 2025).

A further limitation is that many papers use “generative” in different senses. Deep Generative Classifiers for chest X-ray diagnosis insert a stochastic Gaussian latent layer and sample

$$\tilde{\epsilon} = UNet(VAE(x)\oplus\epsilon, t, CLIP(y)),$$9

before multi-label classification (Mao et al., 2018). This improves average AUC modestly across several backbones, but the model does not reconstruct images or define a full generative model over radiographs (Mao et al., 2018). It belongs to the paradigm only in a narrow sense: diagnosis is generated from a latent distribution rather than read off a deterministic embedding. The same caution applies to other hybrid models: not every use of stochasticity or synthesis amounts to a full generative medical world model.

These limitations do not negate the paradigm. Rather, they indicate that rigorous evaluation requires at least three axes: generative fidelity or consistency, diagnostic utility, and reliability of the inferential claim.

7. Broader significance and emerging directions

Across medicine, engineering, and educational assessment, the paradigm supports a shift from rigid prediction to richer forms of inferential structure. In cognitive diagnosis, this shift is explicit: Generative Item Response Theory and Generative NCDM replace transductive latent-parameter estimation with a Generative Diagnosis Function

$$\mathcal{L}_{mse} = \frac{1}{d} \sum_{i=0}^{d-1} (\tilde{\epsilon}_{i} - \epsilon_{i})^2.$$0

which maps response evidence directly to learner and item states (Li et al., 13 Jul 2025). The paper argues that this changes cognitive diagnosis from a transductive to an inductive problem, enabling diagnosis of unseen learners without re-optimization and achieving perfect empirical identifiability scores $$\mathcal{L}_{mse} = \frac{1}{d} \sum_{i=0}^{d-1} (\tilde{\epsilon}_{i} - \epsilon_{i})^2.$$1 and $$\mathcal{L}_{mse} = \frac{1}{d} \sum_{i=0}^{d-1} (\tilde{\epsilon}_{i} - \epsilon_{i})^2.$$2 on both reported datasets (Li et al., 13 Jul 2025). Although this is outside clinical medicine, it is conceptually important: it isolates a general principle of generative diagnosis as state generation from evidence, with response prediction relegated to a downstream reconstruction role.

In industrial diagnosis, the paradigm is also expanding. Syn-Diag uses LLM-based visual-semantic alignment, prompt-conditioned comparative reasoning over fault descriptions, and cloud-edge distillation for few-shot fault diagnosis (Jia et al., 7 Oct 2025). Diagnosis proceeds through semantic grounding and contextual prompt synthesis rather than raw feature-to-label discrimination, while the edge model preserves most cloud performance with roughly 83% parameter and memory reduction and about 50% lower inference latency (Jia et al., 7 Oct 2025). A different direction appears in weak-prior CPS diagnosis, where nominal-data anomaly detection and graph-based root-cause candidate generation produce explanatory diagnosis sets under minimal prior information (Steude et al., 12 Jun 2025). Though not generative in the probabilistic sense, it contributes an abductive mechanism for generating candidate causes from symptoms and causal graphs.

A plausible synthesis is that the generative diagnosis paradigm is converging on four durable principles. First, diagnosis should be treated as inference under partial observation, not merely prediction from a fixed modality contract. Second, generation is valuable when it yields clinically or physically meaningful evidence—balanced subgroups, latent states, hidden modalities, ranked hypotheses, or structural completions. Third, adaptive or case-specific conditioning may become increasingly important, whether through test-time optimization (Zhang et al., 11 May 2026) or instance-specific parameter composition (Tran et al., 29 Mar 2026). Fourth, reliability demands stronger evaluation than image realism alone, including fairness, calibration, post-selection validity, and subgroup robustness (Munia et al., 21 Mar 2025, He et al., 2024).

In that sense, the generative diagnosis paradigm is not a single algorithmic recipe but a reorientation of diagnosis itself. It treats diagnostic systems as models that can generate what is missing, what is latent, what is rare, or what is diagnostically explanatory, and then reason from those generated objects. The paradigm’s most mature forms already show practical gains in fairness-aware dermatology (Munia et al., 21 Mar 2025), virtual sensing for CHF (Na et al., 2023), longitudinal lab diagnosis (Zhang et al., 2017), differential diagnosis (Alam et al., 2023), diagnosis-aware modality synthesis (Shin et al., 2020), and joint generative medical inference (Zhang et al., 11 May 2026). The main unresolved challenge is no longer whether generation can help diagnosis, but under what constraints it can do so reliably, transparently, and robustly enough for high-stakes deployment.