Imputing the Missing Kidney Baseline from Nursing Notes

Most intensive care unit (ICU) models for acute kidney injury (AKI) learn from labels that a rule computes: an episode counts as AKI when blood creatinine rises enough relative to the patient's "normal" level. That normal level, called the baseline, is often never measured. Pipelines then fill it in with a formula that assumes the patient has ordinary muscle mass. Creatinine is a waste product of muscle, so a frail, wasted, or amputee patient has a much lower true baseline than the formula assumes. The same absolute rise is then a smaller relative rise on paper, and real kidney injuries in these patients are labeled as non-events. The result is a training set in which the positive class is too small and missing mostly from one clinically vulnerable subgroup. No class-weighting trick can repair that, because the labels themselves are wrong.

We propose to treat the unmeasured baseline as a missing value and fill it in by multiple imputation, meaning we draw many plausible values rather than one. The imputation model is conditioned on evidence of low muscle mass that exists almost only in free text: nursing descriptions of cachexia, muscle wasting, bed-bound status, or amputation. Each draw of the baseline yields a different set of AKI labels. Downstream models are trained and evaluated across those label sets, so uncertainty about the baseline carries through to the reported results. In two sentences: creatinine-based AKI labels undercount injury in low-muscle patients because the imputed baseline ignores muscle mass. Using the notes to impute that baseline, with uncertainty, should recover the missing positives where rebalancing cannot.

Core Claim

The claim is specific and falsifiable. We restrict to patients whose true pre-admission baseline is known, hide it, and impute it. The text-informed imputation should cut baseline error and AKI label disagreement in the low-muscle subgroup, relative to three alternatives: formula back-calculation, in-stay minimum creatinine, and multiple imputation using structured data only. The gain should be concentrated in that subgroup and near zero elsewhere. If structured proxies alone close the gap, the text component is unnecessary and we will report that.

The mechanism can be written in one line. The Cockcroft-Gault equation estimates creatinine clearance as $(140-\text{age})\cdot w / (72\cdot S)$, where $w$ is weight in kg and $S$ is serum creatinine in mg/dL. Back-calculating a baseline $b$ from an assumed normal clearance therefore scales with $w$. If a patient's true creatinine generation is a fraction $k<1$ of what weight implies, the true baseline is roughly $k\,b$. An observed creatinine $S_t$ then gives an apparent relative rise that is too small by the factor $k$:

$$\frac{S_t}{b} = k \cdot \frac{S_t}{k\,b}.$$

A true 1.5-fold rise, the usual KDIGO stage-1 threshold (KDIGO is the standard clinical guideline defining AKI stages), appears as a $1.5k$-fold rise. With $k=0.7$ that is about 1.05, and the injury disappears from the labels. This is the ICU version of the "creatinine trap" described clinically in Drug Use in Renal Insufficiency, Drug-induced Nephropathy.. That paper describes an 80-year-old woman with creatinine 0.9 mg/dl and clearance of only 54 ml/min. Piccolo Glossario Per Una Migliore Comunicazione Tra MMG e Specialista Nefrologo shows the same trap delaying nephrology referral.

Relation to Prior Work

Three groups of papers set up the problem. The first supplies the label machinery and exposes its fragility. pyAKI automates KDIGO staging and offers several baseline options: minimum, mean, first value, preoperative, and Cockcroft-Gault back-calculation. It fills gaps by carrying values forward. It never asks which baseline is correct for a given patient, and it uses no text. We reuse pyAKI as the labeling engine and replace its single-value baseline with a distribution. An empirical study on KDIGO-defined acute kidney injury prediction in the intensive care unit shows that label definition dominates prevalence: 17% under creatinine-only criteria versus 52% with urine output in HiRID-II. This motivates our focus on labels rather than architectures. Yet Another ICU Benchmark reaches the general version of that conclusion, finding that preprocessing and cohort choices often matter more than the model. It also excludes notes by design, and that exclusion is the gap we target.

The second group shows that notes carry information structured tables lack. Early Prediction of Acute Kidney Injury in Critical Care Setting Using Clinical Notes predicts AKI from first-day notes on MIMIC-III (AUC 0.779, with random undersampling). It removes patients with prior CKD or kidney-related mentions and does not handle negation. Our method needs exactly those removed patients. On the Importance of Clinical Notes in Multi-modal Learning for EHR Data finds that nursing notes, not physician notes, add most of the lift over structured data (decompensation AUPRC 39.7 versus 31.7). That is why our extractor reads nursing notes. Both papers feed text into the predictor. We feed it into the label-generating imputation, which is a different place in the pipeline.

The third group supplies methods and constraints. A Framework for Flexible Extraction of Clinical Event Contextual Properties from Electronic Health Records classifies whether a concept is present, belongs to the patient, and is recent. It handles rare context classes with LLM-generated synthetic examples plus two-phase training, first on balanced data and then on the full distribution. It reports up to 28% higher minority recall with BERT and recall as low as 0.03 for zero-shot Llama 3.1. We reuse this recipe for our own imbalance problem: "patient currently has muscle wasting" is a rare context. The paper never connects extraction to an outcome, and we do. Correlation visualization under missing values shows that imputers with low cell-level error can still distort correlations, while chained-equation imputation (MICE) preserves them. That constrains us to judge imputation by how well it preserves the baseline-to-muscle relationship and the label distribution, not by error alone. It also favors MICE over deep generative imputers as the core engine. Creatinine clearance versus serum creatinine as a risk factor in cardiac surgery and Circulating Adiponectin Is Associated with Renal Function supply outside evidence that demographic equations misstate risk and disagree with each other. They motivate the problem but are not benchmarks for this study.

Research Program

The first component is a contextual muscle-mass extractor for nursing notes. MedCAT (a medical concept recognition tool) finds candidate concepts: cachexia, sarcopenia, muscle wasting, amputation, paraplegia, bed-bound, and poor oral intake. A BERT context classifier, trained with the two-phase learning plus synthetic data recipe, then keeps only mentions that are present, about the patient, and current. "Denies weight loss" and "mother with wasting disease" should not count. We will annotate a few hundred MIMIC-III nursing notes for this purpose, because the CogStack corpus is not something we can assume access to. The output per stay is a small feature vector: counts and flags per concept, plus the earliest time documented.

The second component is the baseline imputation model. It is a chained-equation imputer for log baseline creatinine. Its predictors are age, sex, weight, height, albumin, relevant diagnosis codes, and the text features. It also includes the indicator of whether a prior baseline exists, because missing baselines are not random: patients with no prior labs differ systematically. We fit it on patients with a known prior baseline and draw $m=20$ values for those without one.

For missingness that depends on the unobserved value itself (missing not at random), we run a delta-adjustment sensitivity analysis. This shifts imputed baselines by fixed offsets and reports how far conclusions move.

Each draw is passed through pyAKI, giving 20 label sets. Downstream AKI predictors are trained on each set, and performance and subgroup recall are pooled across them with Rubin's rules, the standard way of combining estimates from multiply imputed data.

Baselines

The naive reference point is pyAKI with Cockcroft-Gault back-calculation for missing baselines and carry-forward within stays. This is the default a benchmark user would pick.

The strongest simple alternative on the baseline question is the in-stay minimum creatinine, also a pyAKI option. It needs no demographic assumption. However, it can be pulled down by fluid dilution and by post-injury recovery, so it may overcount AKI. It is the most plausible competitor because it partly adapts to low-muscle patients automatically.

The strongest principled competitor is MICE with every structured muscle proxy we can assemble and no text: weight, height, BMI, albumin, and diagnosis codes for cachexia, amputation, paralysis, and malnutrition.

The key ablation is identical to our method except that the text features are replaced by keyword counts without context classification. This isolates whether negation and experiencer handling matter.

All imputers are fit on the same training folds, with the same patients, the same $m$, and the same hyperparameter search budget. Beating the structured MICE baseline means a lower median absolute baseline error and higher KDIGO stage agreement with the true-baseline labels in the low-muscle subgroup, with pooled confidence intervals that exclude zero difference. Gains outside that subgroup are not required.

For downstream prediction, we retrain L2 logistic regression on hybrid note features, following Early Prediction of Acute Kidney Injury in Critical Care Setting Using Clinical Notes (reported AUC 0.779 on MIMIC-III), and a gradient-boosted tree model on structured data. We cite the self-correcting RNN of Self-Correcting Recurrent Neural Network for Acute Kidney Injury Prediction in Critical Care (reported AUC 0.893 and F1 0.738 on MIMIC-III) as the strongest reported structured AKI predictor. Its labels follow its own baseline convention, so we will not compare against that number directly. We found no reported result on Lacuna for baseline-creatinine imputation error or for label agreement under masked baselines, so those comparisons are all run within this study.

Why not just…?

Why not just reweight or oversample the AKI class? Reweighting multiplies the weight of the positives you have. The hidden injuries in low-muscle patients are labeled negative, so reweighting pushes the model further toward calling them healthy. It also adds false alarms, the problem noted in An empirical study on KDIGO-defined acute kidney injury prediction in the intensive care unit. We test this directly by applying focal loss and undersampling to the naive labels and measuring recall against true-baseline labels in the low-muscle subgroup. We expect it to stay low, while overall recall rises and precision falls.

Why not just use the structured proxies you already have? This is the real threat. Weight, albumin, and codes for amputation or cachexia may capture most of the muscle signal. The structured MICE baseline exists to settle this. The evidence that it may fall short is indirect. Nursing notes add information beyond tables in On the Importance of Clinical Notes in Multi-modal Learning for EHR Data, and diagnosis codes for frailty are known to be sparsely recorded. If the text features add nothing on top of structured MICE, the contribution reduces to "impute the baseline with MICE instead of back-calculating." That result is still useful, and we would report it as such.

Why not just take the minimum creatinine during the stay? It adapts to each patient without a formula, but it assumes the patient passes through a normal-function state during the admission. It is also biased low by fluid dilution. Masked-baseline evaluation measures this bias directly and separately in patients with and without a large positive fluid balance.

Why not just feed the notes to an end-to-end predictor? A cross-modal model trained on biased labels learns to reproduce the bias. More text in the predictor cannot create positives that the labeling rule deleted. The label-agreement experiment makes this concrete. A note-aware predictor trained on naive labels is scored against true-baseline labels, and its low-muscle subgroup recall is compared with the same predictor trained on imputed labels.

Why not just prompt an LLM to extract frailty? The evidence says zero-shot causal LLMs miss rare contexts, with minority recall around 0.03 in A Framework for Flexible Extraction of Clinical Event Contextual Properties from Electronic Health Records. We include a zero-shot Llama extractor in the extractor comparison anyway.

Proposed Experiments

The decisive experiment uses masked baselines on MIMIC-III. We take adult ICU stays with a creatinine measured 7 to 365 days before admission during an earlier encounter, and we deliberately keep the patients with CKD and kidney-related mentions. We treat that prior value as the true baseline, hide it, and impute it under each method. The primary metrics are three: median absolute error in log baseline, coverage of the imputation intervals, and KDIGO stage agreement (weighted kappa) with the true-baseline labels. Each is reported overall and in the low-muscle subgroup. That subgroup is defined without text, as the lowest quintile of the ratio of known baseline creatinine to weight, so the subgroup definition cannot favor our method. The prediction is a clear drop in subgroup error and a rise in agreement, with little change elsewhere.

A second check uses urine output, whose AKI criterion does not depend on baseline creatinine. Among stays that meet the urine-output criterion, we measure how often each baseline method also flags creatinine-defined AKI. Better baselines should raise this agreement in the low-muscle subgroup. This test also covers patients who have no prior measurement.

A third check addresses the downstream consequences. We retrain the note-based logistic regression and the structured tree model on each label source and pool across imputations. We report AUROC, AUPRC, alarms per patient, and recall on true-baseline positives within the low-muscle subgroup. We expect recall in that subgroup to rise without the precision collapse that reweighting causes.

The MNAR sensitivity analysis reports how far the subgroup prevalence and recall conclusions move under baseline shifts of plus or minus 10 to 20%.

Transfer beyond MIMIC-III is limited because eICU, HiRID, and AUMCdb lack comparable nursing notes. We will therefore test only the structured half of the imputer across centers through YAIB's harmonized cohorts. Any claim about the text component will be confined to MIMIC.