Dynamic Individualized Risk Prediction in IgA Nephropathy

IgA nephropathy is highly variable in its presentation, histopathological pattern of injury, and treatment response. The rate of progression to kidney failure is also highly variable while the lifetime risk of kidney failure is as high as 80%.¹ This variability is a significant source of anxiety for those living with IgA nephropathy,² who are often young, with family and financial responsibilities, and live with the lifetime threat of kidney failure without knowing when this may occur and without tools to provide accurate insights into their evolving risk over time. In this issue, Chen et al. propose a modified recurrent neural network (RNN)-based machine learning model (MLM) to address this problem.³

Over a 100 prognostic biomarkers have been proposed for IgA nephropathy, but only a few have real-world clinical value.⁴ Proteinuria, BP, eGFR, and the histomorphometric features of the Oxford Classification are each widely validated, easily accessible, and affordable to measure.⁴ In isolation, these biomarkers only predict kidney failure well at a population level, performing less well at an individual patient level. The International IgA Nephropathy Prediction Tool (IIgANPT), the current gold standard for IgA nephropathy risk prediction, addressed this issue by combining these clinicopathological biomarkers to produce an individualized percentage risk of 50% decline in eGFR or kidney failure up to 80 months from diagnosis.⁵

While the IIgANPT performs better than individual prognostic markers,⁶ a fundamental shortcoming is that it only predicts risk at a single point in time—the point of biopsy. Although it has been modified and validated for use at subsequent fixed landmark time points (12 and 24 months post biopsy),⁷ it remains unable to dynamically update risk on the basis of changing clinicopathological features over time. There is evidence to indicate value to this approach; time averaged proteinuria and BP, for instance, are better predictors of kidney failure in IgA nephropathy than single time point values.⁴ A tool capable for integrating these data to deliver dynamic risk predictions at any time point would be extremely valuable for patients, clinicians, and researchers. However, development thus far has been limited by a myriad of technical challenges, including the availability and accessibility of sufficiently sophisticated statistical models.

MLMs are advanced computational tools that leverage statistical methods to make data-based predictions. MLMs identify patterns in large datasets during training, which is used to interpret newly provided data to make informed predictions. The effectiveness of MLMs has improved significantly over time because of the growth in computer processing power and the increasingly ubiquitous availability of large datasets for model training. RNNs are a specific type of MLM built to handle sequential data. This is achieved by maintaining an internal state (referred to as a “hidden state”), which is updated with each consecutive input received in sequence. The hidden state acts as the memory of an RNN, capturing relevant information processed from inputs earlier in the sequence. RNNs then integrate the most recent input with the current hidden state of the network to provide an output (a prediction or a decision), which theoretically considers all data points provided earlier in the sequence. These features make RNNs particularly adept for applications in nephrology, where prior sequential outputs (e.g., proteinuria over time) can influence the likelihood of an outcome (e.g., kidney failure).

An inherent limitation of RNNs is a tendency to become less effective or even unstable when processing long data sequences. This issue arises during model training, when gradients calculated with each sequential input are used to update weights assigned to prior data points (influencing the hidden state). With long sequences (e.g., prolonged follow-up data for a slowly progressive disease like IgA nephropathy), these gradients can become exceptionally small or large, leading to the vanishing and exploding gradient problems, respectively. This compromises the model's ability to accurately update its hidden state on the basis of data early in the sequence, significantly reducing effectiveness. Variations of RNNs that allow the accommodation of earlier data points (referred to as “long-term dependencies”) address this limitation, including long short-term memory (LSTM) networks.

LSTMs add a gating mechanism to the architecture of RNNs, which regulates the flow of information within the network. Gates include an input gate, an output gate, and a forget gate, the latter of which selectively removes irrelevant data from the network's hidden state reducing the likelihood of redundant dependencies accumulating. This architecture improves the network's ability to accommodate long-term dependencies, allowing it to train on longer data sequences more effectively than a standard RNN. Although LSTMs require more computational power, they consistently outperform RNNs on tasks involving long-range temporal dependencies.⁸

Chen et al. used an open-source LSTM variant, an interpretable multivariable LSTM (IMV-LSTM), to model individual risk of a 50% reduction in eGFR or kidney failure. Three IMV-LSTM models were trained and internally validated on single-center data collected from 1031 patients over three years—a full, baseline, and time-variant model. The full model was trained on data from a fixed time point (time-invariant data: sex, body mass index, hypertension, and the Oxford Classification) and longitudinal follow-up data (time-variant data: age; proteinuria; mean arterial BP; and serum levels of albumin, creatinine, uric acid, and triglycerides) while the baseline and time-variant models were trained on time-invariant and time-variant variables alone, respectively. When externally validated in a cohort of 1025 patients from 18 centers, the models that incorporated longitudinal data points outperformed the baseline model (c-statistic of 0.93 and 0.92 versus 0.79), as well as the IIgANPT (c-statistic of 0.79).

While the model developed by Chen et al. highlights a value to leveraging longitudinal data and MLMs to enhance risk prediction in IgA nephropathy, several caveats must be considered before their model (or any MLM) can be realistically deployed in routine clinical practice. The first is model generalizability. A robust MLM must be trained on a diverse dataset, which captures a spectrum of clinical variations, to allow reliable pattern recognition when exposed to new data. The model developed by Chen et al. was trained, validated, and tested within a single ethnic group that exhibited limited baseline variability—this increases the risk of model failure among patients with different baseline characteristics, such as minimal proteinuria, severe hypertension, extremes of body mass index, or different ethnicities. This is particularly important for IgA nephropathy, where biopsy features and treatment strategies can vary significantly between countries and ethnicities. A model that is not truly generalizable could also inadvertently propagate systemic health inequalities and, therefore, requires extensive validation before it can be clinically deployed. Training a broadly generalizable model ultimately requires curated international datasets or an acceptance of the model's value being restricted to regional or local populations. Second, while the IIgANPT can be used to estimate risk rapidly in clinical settings using a few accessible clinical data points, which can be manually entered, the model developed by Chen et al. demands more data per patient. Deploying this model in clinical practice will require either substantial data exports for each patient or integration with existing digital health care infrastructures, which may initially limit accessibility. Finally, all MLMs are subject to model drift, requiring regular retraining to refine its predictions and stay current with evolving clinical patterns. This is particularly important for IgA nephropathy, to capture the rapidly changing therapeutic landscape and related effects on long-term outcomes. This necessitates ongoing access to the clinical data entered to model risk and the true outcomes that occurred—this flow of data would need to be supported by a robust digital infrastructure that navigates issues related to cross-border data sharing. Conversely, their model could be adapted for local deployment, allowing retraining to take place locally. This would require appropriate local expertise but may ultimately be more practical.

One of the major criticisms which remain with regard to the application of MLMs to clinical medicine is a lack of explainability. As standalone models, it is fundamentally impossible for users to infer how particular MLMs derive predictions, which carries inherent risks in its applications to clinical medicine. Chen et al. circumvented this issue by using an IMV-LSTM, which identifies variables most weighted by the model to generate predictions. Although the outcomes of this analysis were unsurprising (creatinine, proteinuria, interstitial fibrosis, and BP were among the strongest influencers of the model's decision making), Chen et al. did restrict the variables used to train the model to those broadly known to associate with a poor prognosis in IgA nephropathy. A model trained on a much broader set of clinical variables, perhaps including transcriptomic, proteomic, and pathomic data,⁹ would likely improve prognostic precision and deliver highly personalized information on both prognosis and potential treatment responsiveness, which are urgently needed for IgA nephropathy. The main barrier to this, however, would again be the availability of large uniformly collected international datasets for model training, validation, and testing.

Overall, Chen et al. successfully demonstrate the potential multidimensional deep learning models can bring to clinical nephrology. In the context of IgA nephropathy, dynamic and accurate risk prediction could not only support patient counseling, but also the refinement of individualized treatment strategies and clinical trial recruitment, especially when coupled with approaches to facilitate biomarker discovery. Future development will require international collaboration and data sharing on an unprecedent scale. Fortunately within the global IgA nephropathy community, this is already happening under the auspices of the International IgA Nephropathy Network and the IgA Nephropathy Kidney Atlas.

Disclosures

Disclosure forms, as provided by each author, are available with the online version of the article at http://links.lww.com/CJN/B917.

Funding

None.

Author Contributions

Conceptualization: Haresh Selvaskandan.

Supervision: Jonathan Barratt.

Writing – original draft: Haresh Selvaskandan.

Writing – review & editing: Jonathan Barratt, Haresh Selvaskandan.