Emerging technologies for early risk stratification and precision management of diabetic kidney disease: a multimodal framework integrating digital phenotypes and clinical biomarkers

A 2024 systematic review suggests that polygenic risk scores (PRS) derived from existing GWAS datasets can identify individuals at elevated risk for DKD at the population level (27). However, predictive performance varies with ancestry, phenotype definition, and baseline risk distributions. In real-world settings—particularly those involving mixed ancestry, comorbid cardiovascular disease, or variable medication exposure—model calibration may drift, reducing clinical reliability (28).

Best practices for PRS development and deployment include multi-ancestry training, independent external validation, pre-registered thresholds, and transparent clinical interpretability (27). In this context, PRS may be best utilized not for population-wide screening or as replacements for KDIGO models, but rather as tools for mechanistic stratification or enrichment of clinical trial populations—especially for identifying genetically high-risk individuals with borderline or subclinical phenotypes.

Epigenetic modifications associated with "metabolic memory"—including DNA methylation and histone remodeling—can persist despite glycemic normalization and contribute to podocyte injury, interstitial fibrosis, oxidative stress, and inflammation (29). Recent reviews (2024–2025) have mapped key epigenetic alterations in renal parenchymal cells, including dynamic shifts in DNA methylation and H3K4/H3K27 histone marks (30, 31). Although pharmacologic targeting of epigenetic pathways is promising, robust human data linking these features to hard renal outcomes remain scarce (32).

In parallel, single-cell and spatial multi-omics technologies are rapidly advancing nephrology research (33, 34). Platforms such as scRNA-seq and scATAC-seq allow for precise spatiotemporal mapping of cell-type-specific processes, linking genetic predisposition and epigenetic reprogramming to molecular phenotypes and microenvironmental signals (35).

In the near term, a practical application may lie in multimodal predictive modeling that incorporates epigenetic and transcriptomic features alongside clinical variables, imaging, and fluid biomarkers to refine early risk stratification and subtype identification (34). However, these models require validation in multi-center prospective cohorts with hard endpoints (e.g., eGFR slope, ESKD, cardiorenal events) to establish their incremental predictive value.

Current evidence supports an interactive model in which genetic predisposition, environmental exposures, and epigenetic reprogramming converge to shape DKD trajectory (27, 36). PRS may serve as tools for mechanistic stratification or high-risk trial recruitment, while epigenetic and single-cell approaches guide pathway-based intervention and target prioritization. Successful clinical translation will depend on external multi-ancestry validation, alignment with human kidney tissue data, and demonstration of added predictive utility anchored to meaningful clinical outcomes.

Despite global efforts in assay standardization, serum creatinine remains susceptible to both methodological and interference-related variability. Discrepancies between enzymatic and Jaffe methods, as well as inter-laboratory variation, can result in inconsistent eGFR values and CKD staging errors. External quality assessments and multi-center evaluations continue to detect residual measurement differences even under standardized protocols (40).

Common analytical interferences include falsely low enzymatic readings due to dopamine or dobutamine administration, pseudo-hypocreatininemia caused by IgM paraproteinemia, and interference from N-acetylcysteine. These factors are especially relevant when eGFR values approach clinical decision thresholds (e.g., for drug dosing or nephrology referral) and warrant confirmatory testing or the use of alternative biomarkers when appropriate (41, 42).

The 2021 CKD-EPI equations, which exclude race and incorporate cystatin C, offer improved accuracy and reduced demographic bias compared to traditional creatinine-only models (39). In 2023, a further refinement was proposed: a cystatin C–C-based formula that also excludes sex, offering a robust alternative in patients with abnormal muscle mass, atypical creatinine kinetics, or near dosing thresholds (43).In clinical practice, eGFRcr-cys or cystatin C–only equations should be considered at decision points where misclassification may have therapeutic consequences, provided testing infrastructure allows (14, 38).

When eGFR estimates are discordant with clinical presentation, or in cases involving extreme body composition, abnormal creatinine/cystatin C kinetics, or critical research/drug-dosing scenarios, direct measurement of GFR (mGFR) is warranted. Standardized plasma clearance techniques using non-radioactive tracers (e.g., iohexol) have demonstrated feasibility and reproducibility in both clinical and research settings, and serve as the gold standard for GFR assessment (44).

Like eGFR, uACR measurement requires rigorous standardization (45). Reference methods such as LC-MS/MS and external proficiency testing have revealed systematic discrepancies between routine immunoassays and isotope dilution mass spectrometry (IDMS) (46). Reference materials—such as NIST SRM 2925 (recombinant human albumin)—enable traceability and help reduce inter-laboratory variability, thereby improving comparability in longitudinal studies and multicenter trials (47). Clinical laboratories should disclose the traceability and uncertainty of their assays. In research contexts, preference should be given to externally validated and calibrated platforms.

Robust analytical calibration at the foundational biomarker level is a prerequisite for the meaningful integration of novel molecular, imaging, or digital phenotyping tools. Without such calibration, perceived gains in prediction may reflect the correction of baseline measurement error rather than the true biological signal. Therefore, technical rigor in eGFR and uACR assessment is essential to support the valid clinical deployment of emerging diagnostic technologies.

KidneyIntelX.dkd is a multi-biomarker machine learning–based risk score that integrates TNFR1, TNFR2, and KIM-1 with key clinical variables such as eGFR, uACR, HbA1c, and blood urea nitrogen (56, 57). This model has received De Novo authorization from the U.S. Food and Drug Administration (FDA) for predicting the 5-year risk of progressive kidney function decline in adults with diabetes and early-stage DKD.

The tool supports individualized care pathways by informing follow-up intensity, referral to nephrology, or clinical trial enrollment (58). Both derivation/validation studies and real-world implementations have shown that combining biomarkers with clinical parameters improves discrimination and facilitates actionable risk stratification (56, 58, 59).

Biomarker performance may vary across diabetes subtypes and disease stages (48). For example, in type 1 diabetes, the independent predictive value of urinary KIM-1 for ESKD is attenuated after adjusting for albuminuria, indicating its potential utility in early-stage disease or in conjunction with inflammatory markers (60).

While TNFR1/2 are more robust than many other inflammatory biomarkers, their levels may be affected by systemic inflammation, limiting disease specificity (61). Therefore, clinical interpretation should account for comorbid conditions and be contextualized within the broader phenotypic landscape.

Emerging evidence also suggests that these biomarkers are modifiable by pharmacologic interventions. Randomized trials of sodium–glucose cotransporter 2 (SGLT2) inhibitors have shown reductions in KIM-1 and modest decreases in TNFR1/2, indicating their potential utility for pharmacodynamic monitoring and treatment response assessment (62, 63).

To enable clinical translation of these biomarkers, robust analytical validation, external performance verification, and seamless clinical integration are essential. Assay platforms should clearly define the sample matrix (e.g., serum or plasma), calibration protocols, precision parameters, and dynamic ranges to ensure inter-laboratory comparability (64). Predictive models must undergo external validation across ethnically and geographically diverse cohorts using standardized outcomes such as ≥40% sustained eGFR decline, ESKD, or composite renal/cardiovascular events (65). Reporting should include discrimination (C-statistic, AUC), calibration (slope, Brier score), and clinical utility metrics (e.g., NRI, IDI, decision curve analysis) (66). Integration into routine care should align with the KDIGO 2024 framework, with predefined trigger thresholds for nephrology referral and follow-up intensity. Post-implementation strategies must include recalibration procedures and monitoring for model performance drift (67).

Inflammatory (TNFR1/2) and tubular injury (KIM-1) biomarkers offer independent prognostic value beyond traditional markers and can aid in early identification of high-risk patients, individualized care planning, and enrichment of clinical trial cohorts. However, widespread clinical adoption will require harmonization of assay platforms, validation across diverse populations, and the accumulation of real-world evidence demonstrating improvement in patient reclassification and clinical outcomes.

In a randomized, double-blind, crossover trial, the addition of dapagliflozin for 12 weeks on top of background renin–angiotensin system (RAS) inhibition significantly reduced CKD273 scores and shifted a subset of individuals from a high-risk to a low-risk peptide profile. These findings position CKD273 as a potential pharmacodynamic marker of treatment response (72). Similar proteomic modulation has been observed with empagliflozin, suggesting a class effect of SGLT2 inhibitors on the urinary peptide signature (73). Conversely, the addition of spironolactone in patients with high CKD273 scores failed to prevent microalbuminuria onset, indicating limited evidence for mineralocorticoid receptor antagonist (MRA) therapy guided by this biomarker at present (70). Overall, CKD273 appears more suited as an early "trigger" and a treatment response indicator, rather than a diagnostic surrogate (69).

The development and validation of CKD273 have predominantly relied on CE-MS (74). Despite its strong intra- and inter-assay reproducibility, variability introduced by batch effects, platform heterogeneity (e.g., CE-MS vs. LC-MS/MS), and pre-analytical procedures (e.g., centrifugation, desalting, freeze–thaw cycles) remains a major challenge for multicenter deployment (75, 76). To ensure analytical reliability and comparability, strict quality control, batch normalization, and inter-laboratory calibration protocols must be explicitly implemented in both research and clinical settings.

The CKD273 classifier may serve as a complementary biomarker layer atop standard KDIGO-based risk stratification (e.g., eGFR × uACR, KFRE), with potential applications in three distinct clinical scenarios (69). First, in patients with borderline or fluctuating uACR levels, a positive CKD273 result could trigger enhanced surveillance strategies such as intensified uACR retesting or ambulatory blood pressure monitoring(ABPM). Second, in high-risk individuals without overt proteinuria, CKD273 can help identify subpopulations prone to rapid progression, thereby improving event rates and statistical power in clinical trials. Third, in therapeutic monitoring, a post-treatment reduction in CKD273 score may serve as a pharmacodynamic marker when interpreted alongside uACR decline and eGFR slope stabilization. To ensure clinical relevance, all studies evaluating CKD273 should report not only discrimination metrics (e.g., C-statistic, AUC), but also calibration (e.g., calibration slope, Brier score) and clinical utility indices (e.g., NRI, IDI, decision curve analysis). Furthermore, external validation should prioritize hard clinical endpoints such as sustained ≥40% eGFR decline, ESKD, or cardio–renal composite events to demonstrate incremental prognostic value beyond existing tools.

CKD273 offers incremental risk information before overt albuminuria and exhibits quantifiable responsiveness to SGLT2 inhibitor therapy. Its optimal role is as an early activation signal and a treatment monitoring tool layered on top of standard care. Future directions include large-scale, multi-center prospective studies with harmonized protocols, validation against hard clinical endpoints, and generation of real-world evidence to demonstrate reclassification benefit and net clinical impact.

Although uEVs show potential for early stratification and mechanistic mapping, there remains a lack of large-scale, prospective validation studies using hard renal endpoints such as sustained eGFR decline, ESKD, or cardio–renal composite outcomes (79). Current evidence is primarily derived from cross-sectional or small longitudinal cohorts (80). To assess incremental value, standardized phenotype definitions and methodological consistency are essential.

Technical challenges specific to uEVs include variability in sample dilution, low protein concentrations, and contamination from non-renal sources. Recent international guidelines emphasize the importance of unified nomenclature, standardized isolation techniques such as ultracentrifugation and immunocapture, and normalization approaches using urinary creatinine, volume, or vesicle-specific markers (80). Reporting standards should include particle count, protein content, and characterization markers, alongside inter-batch and cross-platform quality control.

At the current stage of evidence, uEVs are best regarded as a complementary biomarker layer rather than a replacement for conventional tools (80). When used in conjunction with urinary proteomics or serum markers, uEVs may enhance risk stratification within the KDIGO framework. Evaluation should include not only discrimination metrics but also calibration performance, reclassification indices, and clinical decision impact analyses, particularly concerning testing frequency, referral timing, and therapeutic escalation.

Research should prioritize the development of multi-ethnic prospective cohorts, harmonized pre-analytical workflows, and validation strategies anchored to hard clinical endpoints (78). Candidate biomarkers of interest include podocyte-derived proteins and associated microRNAs, as well as vesicular cargo originating from distinct tubular segments or signaling pathways (81, 82). These may support the delineation of tubulointerstitial-dominant phenotypes and enable more personalized treatment approaches. Ultimately, uEVs hold potential as part of an integrated, multi-modal biomarker strategy for early detection, risk classification, and treatment monitoring in DKD.

Recent advances in non-invasive imaging technologies have enabled the direct visualization and quantification of renal tissue alterations that precede detectable changes in conventional biomarkers such as eGFR and uACR (83). Among these modalities, multiparametric magnetic resonance imaging (MRI) has emerged as a robust tool for assessing renal perfusion, microstructure, and oxygenation (84). Techniques such as arterial spin labeling (ASL) quantify cortical blood flow; diffusion-weighted and diffusion tensor imaging (DWI/DTI) evaluate tissue diffusivity and microstructural integrity; and blood oxygen level–dependent (BOLD) MRI captures tissue oxygenation through deoxyhemoglobin signal variation (85). Collectively, these metrics can detect early renal alterations in individuals with diabetes, even before a measurable decline in kidney function (86).

To facilitate clinical translation, initiatives such as PARENCHIMA have proposed standardized acquisition, processing, and reporting protocols across ASL, DWI, BOLD, and T1/T2 mapping (84). These guidelines emphasize the harmonization of pre-scan variables—hydration status, sodium intake, body positioning—and recommend prioritizing cortical regions to improve reproducibility across imaging platforms. Despite these advances, widespread adoption of renal MRI remains limited by high cost, extended scan durations, and inter-vendor variability (87).

In addition to MRI, shear wave elastography (SWE) has been investigated as a non-invasive surrogate for renal fibrosis (88). By quantifying cortical stiffness, SWE offers insight into interstitial fibrosis, supported by biopsy-correlated studies and meta-analyses. Combining SWE with conventional biomarkers (e.g., eGFR) may enhance discrimination of fibrosis stages (88). However, SWE is highly operator- and device-dependent and lacks standardized thresholds across platforms, limiting its broad clinical utility.

These imaging tools are best positioned as adjunctive modalities in selected clinical contexts rather than as general screening instruments. Their integration is particularly valuable in three scenarios (1): enriching clinical trial populations by identifying individuals with greater subclinical tissue injury—such as impaired perfusion, restricted diffusion, or increased cortical stiffness—thereby improving event rates and trial efficiency (2); monitoring therapeutic response in patients receiving renoprotective interventions (e.g., SGLT2 inhibitors, mineralocorticoid receptor antagonists, GLP-1 receptor agonists), where dynamic imaging changes may complement conventional markers like uACR and eGFR slope; and (3) evaluating the incremental prognostic value of imaging biomarkers in prospective cohorts anchored by hard clinical endpoints (e.g., sustained ≥40% eGFR decline, ESKD, or cardio–renal composite outcomes), with emphasis on reclassification performance and net clinical benefit (89).

Multiparametric renal MRI and shear wave elastography (SWE) offer valuable insights into subclinical tissue-level injury and hold promise as precision imaging biomarkers within emerging nephrology frameworks (90). Their optimal role lies in augmenting, rather than replacing, existing clinical assessments, particularly in early detection and treatment response monitoring. Moving forward, priorities include the development of harmonized acquisition protocols, robust multicenter validation across imaging platforms, and demonstration of additive prognostic value over conventional tools such as eGFR and uACR. Additionally, integration with clinical decision pathways will require health-economic evaluations, workflow optimization, and regulatory recognition to support scalable adoption.

Repeated measurements of eGFR and uACR allow the construction of individual trajectories rather than relying on a single value. Models that use these trajectories generally identify high-risk patients earlier and more accurately than static models based only on baseline measurements (96). Incorporating eGFR slope into prediction models has improved calibration for estimating 2-year risk of ESKD. Joint modeling approaches, which analyze biomarker trajectories and time-to-event outcomes together, further strengthen prediction and reduce bias when progression is variable or follow-up is incomplete (97, 98). In parallel, wearable-derived metrics such as CGM-based glycaemic variability and ABPM-based blood-pressure fluctuations provide high-resolution information on exposure to harmful metabolic and hemodynamic stress, and these metrics correlate with adverse cardio-renal outcomes. For these reasons, longitudinal and high-frequency data streams are increasingly viewed as core components of modern DKD risk models (99, 100).

Several time-aware modeling strategies have been applied to DKD. Landmark or dynamic prediction frameworks update an individual's risk estimate at predefined time points based on all information accumulated so far, which mirrors routine follow-up visits in outpatient care (101, 102). Time-dependent Cox models and joint models combine evolving biomarker values with survival outcomes, allowing risk to change as kidney function or albuminuria changes and reducing bias from informative dropout or censoring (103). Deep learning methods, including recurrent neural networks and architectures such as RETAIN and BEHRT, can integrate information from many time points and modalities while retaining some level of interpretability for clinicians (104). Regardless of the specific technique, it is essential to define the observation window (what history is used) and the prediction horizon (what time frame is being predicted), and to train and test models using hard clinical end points to avoid data leakage and to keep the results clinically meaningful (105).

In practice, DKD risk models can combine multiple data types—clinical variables, laboratory tests, omics, imaging markers and digital phenotypes. These can be fused at different stages of the modeling pipeline, such as early fusion (simple concatenation of features) or later fusion of intermediate representations. Real-world datasets often contain irregular sampling and missing values, which require appropriate handling through modality-specific masking, imputation strategies or dropout-based methods. Differences between hospitals, laboratories and devices can also degrade performance when a model is moved to a new setting. Domain-adaptation and federated-learning approaches have been proposed to reduce this problem and to allow multi-center model development without sharing raw patient data (106 –108). A notable example of successful multimodal fusion is KidneyIntelX.dkd, which combines TNFR1, TNFR2, KIM-1 and routine clinical variables and has received FDA De Novo authorization for assessing DKD progression risk, illustrating that such pipelines can reach clinical-grade implementation.

Before deployment, DKD risk models require careful validation. This includes internal validation using approaches such as nested or time-split cross-validation, and external validation in independent cohorts that differ by center, device and patient mix. Key performance metrics include discrimination (AUROC, C-index), calibration (calibration slope, Brier score) and measures of clinical usefulness such as net reclassification improvement, integrated discrimination improvement and decision-curve analysis, ideally reported for different prediction horizons using time-dependent AUROCs (106). Model development and reporting should follow established guidelines such as TRIPOD-AI and PROBAST-AI, and should include fairness assessments across sex, age, ethnic and socioeconomic subgroups. Fairness metrics—for example equalized odds, calibration within subgroups and comparison of error rates—help to ensure that model performance is equitable across diverse populations (109).

After a model is introduced into clinical practice, its performance can change as populations, treatments and measurement practices evolve. Therefore, ongoing monitoring is required to detect data drift, loss of calibration and deterioration in specific subgroups. Monitoring can use simple tools such as stability indices, calibration plots and periodic assessment of net clinical benefit. When pre-specified thresholds are crossed, a stepwise recalibration plan—starting with simple updates to the intercept and slope and escalating to partial or full retraining—should be implemented (110). Governance frameworks, including the FDA's Good Machine Learning Practice (GMLP) and the NIST AI Risk Management Framework (AI RMF), provide guidance on documentation, audit trails and change-control processes that support safe model updating and lifecycle management (111). Repeated fairness audits and real-world outcome evaluations are also important to maintain trustworthiness over time (112).

In multicenter CKD cohorts, joint models that incorporate eGFR slope together with time-to-event data have improved both calibration and discrimination for predicting kidney failure compared with simpler approaches (91). KDIGO guidelines support the use of ABPM for CKD risk stratification, and CGM is increasingly used in patients with advanced CKD and dialysis for time-in-range–based glycaemic management, providing rich inputs for risk models. KidneyIntelX.dkd is an example of a regulatory-grade model that combines biomarker and clinical data and has been evaluated in prospective and real-world settings, offering a practical template for future translational efforts (113). These examples show that clinically relevant, externally validated and equity-aware modeling frameworks for DKD risk stratification are feasible in real-world practice.

Overall, machine-learning models for DKD risk prediction should be built on longitudinal data, integrate multiple complementary data sources and be trained on clearly defined, clinically meaningful endpoints. Only models that undergo rigorous validation, fairness assessment and ongoing performance monitoring are likely to deliver sustained benefit in everyday care. Aligning these efforts with clinical guidelines such as KDIGO and KFRE, and with regulatory frameworks such as GMLP and AI RMF, is crucial for technical robustness, clinician confidence and wider adoption across diverse healthcare systems (114).

Traditional Chinese Medicine (TCM) emphasizes zheng (syndrome) as a reflection of dynamic physiological imbalance (115). In the context of DKD, zheng can be reframed as a quantifiable digital phenotype that complements modern risk stratification (116). Advances in sensor and imaging technologies enable the digitization of classical TCM diagnostics: tongue imaging (inspection), pulse waveform acquisition (palpation), structured symptom questionnaires (inquiry), and potentially voice or breath sensing (listening/smelling) (117). Among these, tongue imaging and pulse signals are the most developed. Standardized tongue images yield features such as coating thickness, texture, and color, associated with inflammation, microcirculation, and metabolic status (118). Pulse waveforms captured via photoplethysmography (PPG) offer non-invasive measures of vascular compliance and autonomic tone (119).

To ensure multi-site generalizability, strict acquisition protocols are essential. For tongue imaging, a consistent light source, shooting angle, and exclusion of blurry or overexposed images are required (120, 121). Pulse signals must be acquired in quiet, temperature-controlled environments, with signal quality assurance. Device heterogeneity and environmental variability necessitate pre-specified normalization, calibration pipelines, and population-specific validation. These steps are crucial to establishing reliability before integration into predictive models or clinical workflows (122).

Digital phenotypes can enhance early DKD risk identification by detecting subclinical changes preceding proteinuria or eGFR decline (123). Their low cost, self-collectibility, and suitability for high-frequency monitoring make them valuable for continuous risk surveillance. When used alongside CGM, ABPM, laboratory indices, and imaging markers, these features may enrich predictive accuracy and personalize follow-up intensity. They also provide actionable signals in remote or resource-limited settings, enabling early alerts and behavioral interventions.

TCM-derived digital features can be incorporated into multimodal machine learning models targeting hard renal endpoints such as sustained eGFR decline or dialysis initiation (123, 124). Studies using tongue image datasets have shown predictive utility for metabolic traits, while pulse waveform analyses correlate with arterial stiffness—a known predictor of renal outcomes (125, 126). These signals should be evaluated not only by AUROC but also by calibration, reclassification, and clinical decision utility. Promising use cases include screening, remote monitoring, and adjunctive response evaluation. Their deployment must ensure external validation, interpretability, and fairness across subpopulations.

While TCM-derived digital phenotypes offer an attractive, low-cost avenue for continuous monitoring, their current evidence base is substantially weaker than that of established digital health tools such as pulse wave analysis, ABPM, and continuous glucose monitoring (CGM). Most tongue and pulse studies are single-center, cross-sectional, or small longitudinal cohorts with heterogeneous inclusion criteria and limited adjustment for confounders. Acquisition protocols vary widely across devices and environments, which can introduce substantial measurement noise and reduce cross-site reproducibility.

Moreover, cross-cultural generalizability remains a key challenge. Tongue color, coating, and pulse characteristics can be affected by diet, oral hygiene, skin tone, and comorbid conditions, all of which may differ across populations. In contrast, CGM and ABPM have undergone large-scale, randomized and observational validation, are incorporated into international guidelines, and have well-defined clinical thresholds. At present, therefore, TCM-derived digital phenotypes should be regarded as exploratory signals that may enrich multimodal models or support remote hypothesis generation, but they cannot replace validated digital tools for risk stratification or treatment guidance.

Factors such as skin tone, oral hygiene, comorbidities, and acquisition environment can introduce bias. Models must report performance stratified by age, sex, ethnicity, and device type. Fairness-aware modeling strategies (e.g., reweighting, subgroup-specific training) are critical to mitigate disparities (120). Data privacy, auditability, and version control should be embedded through localized or federated learning pipelines. Ultimately, the value of these digital phenotypes will be determined by prospective validation, cost-effectiveness, and their ability to accelerate time-to-intervention within DKD care pathways.