Nephrology-Oriented Representation Learning (NORA)

• NORA is a nephrology-focused representation learning framework that organizes latent spaces and supervision around kidney-specific clinical problems across various modalities.
• The adult CKD pipeline uses supervised contrastive learning combined with a nonlinear Random Forest on non-renal EHR features to detect early-stage CKD when key biomarkers are missing.
• Multimodal extensions of NORA are applied to pediatric temporal monitoring, renal oncology vision-language tasks, and kidney single-cell analysis, demonstrating broad clinical impact.

Searching arXiv for the cited NORA and related nephrology representation-learning papers to ground the article. arXiv search query: (Khan et al., 16 Sep 2025) Nephrology-Oriented Representation leArning (NORA) denotes a nephrology-focused representation-learning strategy in which the latent space, supervision, and downstream tasks are organized around kidney-specific clinical problems rather than general-purpose medical prediction. In the available literature, the acronym is introduced for chronic kidney disease (CKD) classification from tabular outpatient electronic health record (EHR) data and is then used more broadly as a conceptual blueprint for pediatric temporal modeling, disease-centric vision-language learning in renal oncology, and kidney-focused single-cell foundation modeling (Khan et al., 16 Sep 2025, Durica et al., 17 Nov 2025, Tao et al., 22 Aug 2025, Li et al., 30 Sep 2025). Across these instantiations, the common objective is to learn representations that preserve nephrology-relevant structure under modality-specific constraints such as absent renal laboratory biomarkers, irregular longitudinal sampling, multimodal fusion, or cross-species molecular heterogeneity.

1. Conceptual scope and terminology

NORA is not a single architecture. In adult outpatient CKD classification it is a two-stage pipeline combining supervised contrastive learning with a nonlinear Random Forest classifier on tabular EHR features (Khan et al., 16 Sep 2025). In pediatric nephrology it appears as a temporally aware representation-learning blueprint built around recurrent modeling of longitudinal laboratory sequences and demographic covariates (Durica et al., 17 Nov 2025). In renal oncology it is defined as a disease-centric paradigm for vision-language foundation models whose embeddings are aligned with the kidney care workflow, from anatomical characterization to prognosis (Tao et al., 22 Aug 2025). In kidney single-cell biology, the term is not explicitly used, but the work is presented as conceptually and practically exemplifying kidney-focused representation learning through organ-specific pretraining and evaluation (Li et al., 30 Sep 2025).

Study	Data and modality	NORA instantiation
(Khan et al., 16 Sep 2025)	Tabular outpatient EHR, non-renal variables	Supervised contrastive encoder + Random Forest
(Durica et al., 17 Nov 2025)	Pediatric longitudinal labs + demographics	GRU-based temporal embedding
(Tao et al., 22 Aug 2025)	CT scans + radiology reports	Disease-centric vision-language foundation model
(Li et al., 30 Sep 2025)	Single-cell, single-nucleus, ATAC, spatial omics	Kidney-focused transformer foundation model

A common misconception is to equate NORA with a single CKD classifier or with renal-biomarker-driven modeling. The existing literature shows the opposite. The flagship adult study is explicitly motivated by settings in which serum creatinine and blood urea nitrogen are unavailable or delayed, whereas later work extends the nephrology-oriented representation-learning idea to imaging, temporal pediatric monitoring, and molecular biology (Khan et al., 16 Sep 2025, Tao et al., 22 Aug 2025).

2. Adult outpatient CKD classification: the flagship formulation

The adult NORA formulation addresses a specific outpatient problem: early CKD detection when laboratory measurements required for estimated glomerular filtration rate (eGFR) staging are missing. The clinical staging referenced in the study partitions patients into CKD Stages 1–5 based on common nephrology thresholds: Stages 1–2 for eGFR $\ge 60$ mL/min/1.73m$^2$, Stage 3a for $45\text{–}59$, Stage 3b for $30\text{–}44$, Stage 4 for $15\text{–}29$, and Stage 5 for $<15$, with End-Stage Renal Disease (ESRD) reflecting dialysis dependence or post-transplant care (Khan et al., 16 Sep 2025). Rather than using these renal laboratory values as predictors, the model uses routinely collected non-renal variables: sociodemographic factors, anthropometrics, comorbidities, and urinalysis findings such as proteinuria and hematuria.

The primary evaluation cohort is a de-identified clinic-based EHR dataset from Riverside Nephrology Physicians, a community nephrology practice in Southern California. It contains 960 adult patient records from 2014–2025 labeled with CKD Stages 1–5 and ESRD. For modeling, nephrologist-guided label collapsing yields a binary task: early-stage CKD, defined as Stages 1–2, versus moderate-to-advanced CKD, defined as Stages 3–5 plus ESRD. This produces 183 Class 0 and 777 Class 1 patients, a substantial class imbalance. The cohort has mean age $70.0 \pm 14.0$ years, balanced sex distribution at 50.3% female, and racial composition of Hispanic 43.6%, White 37.2%, Black 13.8%, and Asian 5.4%. Stage counts are Stage 1: 58, Stage 2: 125, Stage 3: 374, Stage 4: 226, Stage 5: 48, and ESRD: 129. Hypertension is present in 74.1%, diabetes mellitus in 37.0%, diabetic nephropathy in 18.6%, proteinuria in 18.1%, hematuria in 0.8%, and dyslipidemia in 9.4%, with modest missingness for height and weight at 5.4% each (Khan et al., 16 Sep 2025).

Class-stratified analysis yields clinically interpretable non-renal signals. Class 0 patients are significantly younger with $p < 0.0001$, and proteinuria is more prevalent in Class 0 than in Class 1 at 26.2% versus 16.2% with $p = 0.0027$. Class 1 has higher hypertension at 77.5% versus 59.6% with $p < 0.0001$, and higher diabetes at 38.2% versus 31.7% with $^2$0. Hispanic individuals are overrepresented in Class 0, whereas Asians are more frequent in Class 1. The study interprets these patterns as mirroring known epidemiology and as providing clinically meaningful non-renal signals for CKD risk stratification (Khan et al., 16 Sep 2025).

Preprocessing follows a conventional but tightly specified tabular EHR pipeline. Numeric features such as age, height, and weight are mean-imputed and standardized with z-score normalization. Sex and race are one-hot encoded, while binary comorbidities are label-encoded as 0/1. Stratified sampling preserves class proportions across train/test splits. In training folds only, SMOTE+Tomek is applied to synthesize minority samples and remove Tomek links, producing a balanced training set of 536 versus 536. Five-fold stratified cross-validation is used, and no explicit feature selection is performed (Khan et al., 16 Sep 2025).

3. Algorithmic structure of NORA

The adult NORA pipeline has two stages: supervised contrastive representation learning on tabular EHR features, followed by training of a nonlinear Random Forest classifier on frozen encoder embeddings (Khan et al., 16 Sep 2025). The data are written as

$^2$1

with $^2$2 corresponding to CKD Stages $^2$3 and $^2$4 to Stages $^2$5, including ESRD.

The encoder $^2$6 is a three-layer multilayer perceptron with hidden dimensions $^2$7 and ReLU activations. The projection head $^2$8 is a two-layer multilayer perceptron mapping to a 32-dimensional embedding, and the resulting vectors are $^2$9-normalized. There are no data augmentations for the tabular inputs; supervision enters through batch-wise label information. For a batch of size $45\text{–}59$0, the supervised contrastive objective follows Khosla et al. and is written as

$45\text{–}59$1

where $45\text{–}59$2, the temperature is fixed at $45\text{–}59$3, positives share the same class label, and similarity is the dot product after normalization (Khan et al., 16 Sep 2025).

Training of the encoder and projection head is joint, using Adam with learning rate $45\text{–}59$4, batch size 64, and 100 epochs. After pretraining, the projection head is removed and the encoder is frozen. Each patient is represented by $45\text{–}59$5, a 32-dimensional embedding passed to a Random Forest classifier with $45\text{–}59$6 and $45\text{–}59$7, with other hyperparameters left at defaults (Khan et al., 16 Sep 2025).

The rationale for the Random Forest is explicitly comparative. A logistic regression classifier on learned embeddings, denoted SCL+LR, underperforms, which the study interprets as evidence that the latent geometry remains nonlinearly separable because of complex feature interactions and residual overlap even after SMOTE+Tomek. The downstream objective is written for clarity in terms of binary cross-entropy,

$45\text{–}59$8

although the paper notes that standard Random Forest fitting uses impurity-based splits. No threshold tuning, no probability calibration, and no explicit early stopping criteria are reported (Khan et al., 16 Sep 2025).

Qualitative evaluation is integral to the representation-learning claim. The paper uses t-SNE visualizations of the supervised-contrastive space and decision-boundary plots contrasting logistic regression with Random Forests, showing more distinct Class 0 versus Class 1 clusters after supervised contrastive learning and better boundary fitting by the nonlinear classifier (Khan et al., 16 Sep 2025).

4. Empirical results and interpretation

On the Riverside Nephrology Physicians cohort, using only non-renal variables, NORA achieves Class 0 F1 = 76.0 and Class 1 F1 = 88.0, with macro-average F1 = 82.0, weighted-average F1 = 84.2, and accuracy = 84.4 (Khan et al., 16 Sep 2025). This exceeds several strong baselines trained on raw features, including XGBoost with Class 0 F1 = 74.0, Class 1 F1 = 88.0, and accuracy = 83.4; LightGBM with the same class-wise F1 values and accuracy = 83.0; and Random Forest on raw features with Class 0 F1 = 70.0, Class 1 F1 = 87.0, and accuracy = 81.6. Relative to SCL+LR, the NORA pipeline raises early-stage CKD F1 from 69.0 to 76.0 and advanced-stage CKD F1 from 84.0 to 88.0. ROC analysis shows AUC 0.888, slightly above LightGBM at 0.886 and XGBoost at 0.885.

On the UCI CKD dataset, which contains both non-renal variables and renal biomarkers, all methods perform strongly because of informative laboratory features and only mild imbalance. NORA attains Precision 98.6, Recall 98.5, F1 = 98.5, and Accuracy = 98.0, competitive with the best baselines such as LightGBM with F1 = 98.5 and Accuracy = 98.0 (Khan et al., 16 Sep 2025). The paper is explicit that this setting mainly verifies cross-cohort robustness; its main impact remains the lab-agnostic outpatient setting.

Interpretability in the adult NORA paper is cohort-level rather than model-specific. Feature importance analyses such as Random Forest importances or SHAP are not reported. Instead, clinical insight is extracted from stratified cohort differences: hypertension and diabetes are more prevalent in advanced CKD, while proteinuria is more common in early-stage CKD within this cohort, consistent with the observation that urinary protein can be an early marker of renal injury (Khan et al., 16 Sep 2025). A plausible implication is that the learned embedding space is leveraging clinically meaningful proxies of renal dysfunction rather than merely exploiting spurious demographic structure, although the absence of explicit fairness analyses means that such an inference remains incomplete.

5. Longitudinal antecedents and pediatric extensions

Earlier nephrology-oriented representation learning predates the NORA acronym. In diabetic patients at the University of Virginia Health System, the sparse longitudinal representations framework models CKD onset using diagnosis-code histories grouped into 283 AHRQ Clinical Classification Software categories over a 12-month observation window with a 0.5-year hold-off period (Zhang et al., 2020). The core method, Sparse Longitudinal Representations (SLR), uses logistic loss with sparse group lasso regularization to enforce global feature sparsity and temporal group-wise shrinkage, while Weighted Bagging of SLR (WB-SLR) adds out-of-bag weighting over bootstrap models. In experiments on a strict cohort of 395 diabetic individuals, WB-SLR reaches AUC $45\text{–}59$9, Sens $30\text{–}44$0, Spec $30\text{–}44$1, and F2 $30\text{–}44$2, outperforming AFV, BPS, ATV, and simple bagged SLR. Interpretability arises directly from sparse coefficients indexed by diagnosis group and time window, highlighting later diabetes with complications, anemia, hypertension, cardiovascular disease, and toxicological abnormalities as positive signals for impending CKD (Zhang et al., 2020).

A distinct antecedent is the Time-Aware LSTM autoencoder for irregularly sampled CKD trajectories. Using stage 3 CKD cohorts from DARTNet and MIMIC-III, the model encodes eGFR time series with explicit time gaps $30\text{–}44$3 and reconstructs sequences in reverse chronological order (Luong et al., 2018). The paper identifies a methodological issue in prior T-LSTM autoencoder usage: representing a patient only with the final hidden state $30\text{–}44$4 discards information in the final memory state $30\text{–}44$5. Using the concatenated representation $30\text{–}44$6 improves separability in synthetic experiments. For slope-differentiated clusters, the average Silhouette coefficient rises from 0.0548 with hidden-only embeddings to 0.4906 with memory-only embeddings and 0.4858 with hidden+memory embeddings; for intercept-differentiated clusters, hidden-only embeddings remain strongest at 0.6169, but the study shows that memory-based embeddings better organize noise structure (Luong et al., 2018). This suggests that nephrology-oriented representation learning has long depended on clinically aligned treatment of time, irregular sampling, and latent trend structure.

The pediatric extension at Great Ormond Street Hospital moves NORA toward temporally indexed monitoring rather than static CKD staging. The cohort comprises 826 children from 2019–2025, selected from an initial extraction of 1,032 patients with at least one nephrology episode, with inclusion requiring at least three creatinine measurements on separate days before a fixed 30-day prediction window and, for deceased patients, at least one creatinine measurement within the final 30 days of life (Durica et al., 17 Nov 2025). Fifteen clinically relevant laboratory markers are modeled, though the analytes are not enumerated; each laboratory test at each event time is encoded through two binary indicators, presence and abnormality. Sequences are anchored to days with valid creatinine measurements, padded to length 100, and processed by a GRU encoder whose final hidden state is concatenated with age and sex and passed to a linear classifier with sigmoid output. The train–validation–test split is stratified 70/10/20, and discrimination is evaluated by AUROC with 95% confidence intervals estimated by 2,000 bootstrap resamples. The paper reports “good discrimination” but no numeric AUROC. It also notes partial separation of GRU-derived patient embeddings under t-SNE, and it frames the work as a practical, temporally aware NORA blueprint tailored to pediatric care (Durica et al., 17 Nov 2025).

6. Multimodal and foundation-model generalizations

The broadest published uses of NORA shift from CKD classification to large-scale multimodal and molecular representation learning. In renal oncology, RenalCLIP is presented as a disease-centric vision-language foundation model trained on 27,866 preoperative CT scans from 8,809 patients across Chinese medical centers and TCIA, with two-stage pretraining that first enhances the image and text encoders with domain-specific knowledge and then aligns them through a CLIP-style contrastive objective (Tao et al., 22 Aug 2025). The image encoder is a 3D ResNet-18 for volumetric CT, while the text encoder is Llama3 8B transformed via LLM2Vec. RenalCLIP spans 10 core tasks covering R.E.N.A.L. nephrometry components, benign-versus-malignant classification, indolent-versus-aggressive renal cell carcinoma classification, and survival prediction. Reported performance includes recurrence-free survival C-index 0.726 in TCIA, malignancy classification AUC 0.841 and F1 0.876 on the combined external cohort, text-to-image mean Recall@5 of 0.120 across five cohorts, and superior report generation metrics relative to RadFM, CT-CHAT, GPT-4o, and MedGemma (Tao et al., 22 Aug 2025). In this formulation, NORA means organ- and disease-specific pretraining, cross-modal image-text alignment, and multi-task utility across the kidney cancer workflow.

At the cellular scale, Nephrobase Cell+ extends the same nephrology-oriented logic to kidney genomics. The model is pretrained on approximately 100 billion tokens from approximately 39.5 million single-cell and single-nucleus profiles across 4,319 samples, spanning human, mouse, rat, and pig, with scRNA-seq, snRNA-seq, snATAC-seq, and spatial transcriptomics modalities (Li et al., 30 Sep 2025). Its architecture is a transformer encoder-decoder with gene-token cross-attention and a mixture-of-experts module, trained with masked generative pretraining under a Zero-Inflated Negative Binomial reconstruction loss, focal cell-type classification, Elastic Cell Similarity regularization, supervised contrastive loss, and adversarial batch and assay removal. In held-out human kidney data, the 1B model reports ARI = 0.82, NMI = 0.78, silhouette $30\text{–}44$7, cLISI = 1.00, iLISI $30\text{–}44$8, kBET $30\text{–}44$9, and Total integration score 0.71; cross-species evaluation reports NMI up to 0.75, ARI up to 0.72, cLISI = 1.00, and $15\text{–}29$0 zero-shot annotation accuracy for major nephron lineages (Li et al., 30 Sep 2025). The paper explicitly argues that organ-centric training captures nephron segment programs, microenvironmental crosstalk, and pathology-associated states better than generalized models.

Taken together, these multimodal studies indicate that NORA has become a broader design doctrine: nephrology-specific corpora, kidney-aligned supervision, and evaluation protocols that privilege clinical or biological tasks native to the renal domain rather than generic transfer benchmarks (Tao et al., 22 Aug 2025, Li et al., 30 Sep 2025).

7. Limitations, unresolved issues, and future directions

Across the literature, NORA systems remain constrained by dataset scope, evaluation gaps, and deployment concerns. The flagship adult CKD study is single-center and adult-only, may encode local practice patterns and demographic biases, does not address stage-wise multiclass classification, and does not evaluate fairness across demographic subgroups or calibrate predicted probabilities (Khan et al., 16 Sep 2025). It also reports Accuracy, Precision, Recall, F1-score, and ROC-AUC, but not PR-AUC, confidence intervals, or statistical tests on classifier outputs. The pediatric pilot likewise omits baseline comparators, does not report a numeric AUROC, handles irregular sampling only implicitly through event indexing, and leaves subgroup performance, calibration, and external validation as future work (Durica et al., 17 Nov 2025). Earlier code-based longitudinal CKD prediction inherits label noise from diagnosis-code definitions and lacks external validation, while the T-LSTM CKD autoencoder remains univariate in eGFR and does not compare directly against standard LSTM or GRU baselines (Zhang et al., 2020, Luong et al., 2018).

The multimodal foundation-model variants introduce different limitations. RenalCLIP is retrospective, predominantly trained on Chinese data, and still requires prospective validation; rare renal tumor subtypes and radiogenomic endpoints remain open areas, and report generation can exhibit factual inconsistencies that necessitate human oversight (Tao et al., 22 Aug 2025). Nephrobase Cell+ is computationally expensive, trained on a fixed 32,768-gene ortholog space, and may underrepresent rare cell types or extreme pathologic states; its embeddings remain correlational and require experimental validation for mechanistic claims (Li et al., 30 Sep 2025).

The forward trajectory described across these papers is consistent. Adult outpatient NORA proposes few-shot and self-supervised learning under scarce data conditions, multimodal expansion to vitals time series and imaging, and prospective validation (Khan et al., 16 Sep 2025). Pediatric NORA calls for time-delta-aware models, richer renal outcomes such as AKI staging and CKD trajectories, and clinically actionable thresholding supported by calibration and decision-curve analysis (Durica et al., 17 Nov 2025). The oncology and single-cell branches point toward a nephrology ecosystem of disease-centric foundation models in which organ-specific pretraining, contrastive alignment, and kidney-native evaluation become the default strategy for representation learning across tabular EHRs, longitudinal monitoring, imaging, and omics (Tao et al., 22 Aug 2025, Li et al., 30 Sep 2025).