Optimizing anemia management using artificial intelligence for patients undergoing hemodialysis

Ethics assertion

This research adhered to the rules of the Declaration of Helsinki and was authorised by the Institutional Evaluate Board of Kangwon Nationwide College Hospital (KNUH-2023-01-011), which waived the requirement for acquiring knowledgeable consent as a result of retrospective nature of the observational research of medical information. This research was retrospectively registered.

Knowledge assortment

We enrolled patients aged ≥ 18 years undergoing hemodialysis in Kangwon Nationwide College Hospital between 2017 and 2022. Throughout this era, 5 nephrology specialists prescribed medicines, together with ESAs. The patients analyzed on this research obtained solely darbepoetin-alfa amongst a number of ESA sorts. As well as, information have been retrospectively collected from inpatient and outpatient clinics.

Info concerning the next scientific variables was obtained: demographic traits, dialysis, drug use, laboratory assessments, and transfusion. The demographic information included intercourse, age, and complete dialysis development date. Dialysis information encompassed systolic blood strain, diastolic blood strain, and pulse price, measured each hour throughout dialysis, together with the dry weight and complete ultrafiltration quantity. The drug classes included ESAs, oral iron, and intravenous iron. Laboratory take a look at information included: Hb degree; white blood cell depend; platelet depend; and ranges of blood urea nitrogen, creatinine, sodium, potassium, chloride, uric acid, complete calcium, phosphorus, intact parathyroid hormone, glucose, complete protein, albumin, aspartate aminotransferase, alanine aminotransferase, alkaline phosphatase, complete ldl cholesterol, triglyceride, high-density lipoprotein, iron, complete iron-binding capability, ferritin, hepatitis B floor antigen, hepatitis B floor antibody, and hepatitis C antibody. Lastly, within the transfusion class, we solely collected RBC transfusion information.

Knowledge preprocessing

Demographic, dialysis, drug use, laboratory take a look at, and transfusion information have been initially collected from 829 patients. Following information assortment, meticulous information refinement was performed primarily based on the inclusion and exclusion standards. Patients with < 3 months of information have been excluded to make sure a complete understanding of particular person scientific adjustments over time. Moreover, solely patients who underwent dialysis periods from 3 h 30 min to 4 h 30 min, reflecting typical dialysis situations, have been included within the dataset. Finally, a dataset comprising 252 patients was constructed, as proven in Fig. 1.

The ESA dose used for coaching and validating the ESA dose advice mannequin and the transfusion alarm mannequin is calculated as the full dose prescribed for the given month. In different phrases, the ESA dose is handled as a single variable. For example, the ESA dose for Could in a affected person who undergoes hemoglobin testing on the thirteenth of every month is the full quantity of ESA prescribed from April 13 to Could 12. The ESA dose on the prescription date is excluded since blood attracts for hemoglobin testing are usually performed earlier than the ESA prescription.

The dataset used for coaching is organized primarily based on patient-months, which correspond to the dates when patients endure hemoglobin assessments. Sometimes, hemoglobin assessments are performed as soon as a month. Since not all patients are examined on the identical date, the month-to-month information is structured in response to every affected person’s particular person take a look at date. In some circumstances, if a affected person’s situation worsens or there’s a particular purpose, an extra take a look at could also be carried out earlier than a whole month has handed because the earlier take a look at. The extra take a look at outcomes are excluded in such circumstances, and solely the common hemoglobin take a look at outcomes are used.

Variables with lacking values exceeding 15% of the complete dataset have been eliminated, whereas these with lacking values beneath 15% have been addressed by imputation using previous values from the identical affected person. A label-encoding methodology was employed for categorical information, facilitating a streamlined evaluation throughout constant classes.

Mannequin structure

We launched GAM to successfully seize historic info from the info of patients undergoing persistent hemodialysis whereas concurrently specializing in crucial info inside a comparatively brief timeframe. The GRU¹⁸ displays computational effectivity and effectiveness in dealing with time-series information. Moreover, the mixing of a multi-head consideration mechanism¹⁹ empowers the discernment of interrelationships among the many information inside the identical window.

The general structure of GAM is depicted in Fig. 2. Initially, affected person information, together with demographics, dialysis, drug use, laboratory assessments, and transfusions, have been fed into the GRU and a focus modules. Affected person information was entered into the GRU items sequentially, leading to hidden states cascading to the following GRU items, thereby capturing sequential info. Concurrently, affected person information have been inputted into the eye module to extract important particulars inside the present window. The extracted info was then concatenated to kind the ultimate affected person embedding that encapsulated the affected person’s historic information. On this context, (:xleft(0right)), (:t), and (:h) denote the primary information level inside the present window, window measurement, and hidden state, respectively, with the preliminary hidden state set to a random worth.

The generated affected person embedding encapsulates wealthy historic info on patients undergoing persistent hemodialysis. The method of leveraging this embedding to forecast Hb ranges is illustrated in Fig. 3. Just like the beforehand described methodology, GAM makes use of time-series information encompassing Hb ranges, ESA doses, and different supplementary information from time factors t to t + 3 to generate the affected person embedding. Subsequently, the resultant embedding is fed into the enter of the Hb predictor, which consists of totally linked and dropout layers, in the end predicting the Hb ranges at time level t + 4.

ESA dose advice and transfusion alarm framework

The general framework for conducting the ESA dose advice and transfusion alarm duties is illustrated in Fig. 4. In contrast to the Hb prediction activity, affected person information starting from time factors t to t + 2 have been entered into the GAM community to assemble the affected person embedding. The ensuing embedding was mixed with affected person information at time level t + 3, encompassing Hb ranges, transfusion quantity, and supplementary information. Importantly, the optimum Hb degree at time level t + 4 was additionally included at this stage.

Our main objective was to keep up the affected person’s Hb degree inside the regular vary by using the ESA dose advice and transfusion alarm. To realize this, we introduce an method whereby, as an alternative of using the precise t + 4 Hb degree, we inputted details about the best Hb degree that may be adjusted by ESA administration and transfusion. This method enabled the mannequin to appropriately advocate the ESA dose and determine when transfusion is required. The optimum Hb degree corresponded to the identical time level as that for the ESA dose advice and transfusion alarm (t + 4). The optimum Hb degree was adjusted primarily based on the Hb degree at time level t + 3, as described in Eq. 1⁹:

The notation (:{widehat{y}}_{hb}^{t+4}) represents the optimum Hb degree at time level t + 4. If the Hb degree was < 10, (:{widehat{y}}_{hb}^{t+4}) was set as (:{y}_{hb}^{t+3}) + 1. If the Hb degree exceeded 12, (:{widehat{y}}_{hb}^{t+4}) was calculated by subtracting 1 from (:{y}_{hb}^{t+3}). For circumstances outdoors these ranges, (:{widehat{y}}_{hb}^{t+4}) was set to 11. Based on the Kidney Illness Outcomes High quality Initiative (KDOQI) guideline²⁰, the optimum hemoglobin degree is 11 g/dL or extra however doesn’t exceed 13 g/dL, whereas the Kidney Illness Bettering World Outcomes (KDIGO) pointers advocate a degree of 10–11.5 g/dL⁶. Provided that insurance coverage covers ESAs solely as much as a Hb degree of 11 g/dL in South Korea, this method was established. Subsequently, affected person embedding was concatenated with the affected person information at time level t + 3 and the optimum Hb degree at time level t + 4. This concatenated info was then processed by totally linked and dropout layers.

The ESA dose suggestions have been categorized into three teams: extra, related, and fewer, indicating whether or not the next, related, or decrease ESA dose ought to be prescribed for the following month (t + 3) in comparison with that within the earlier month (t + 2). You will need to observe that every affected person’s response to ESA varies; subsequently, it may be delicate to even small adjustments within the ESA dose, and the impact could also be sluggish even when a comparatively great amount of ESA is run. To accommodate such variations, the “related” class thought of doses inside ± 25% of the ESA dose prescribed within the earlier month (t + 2). The transfusion alarm outputs “mandatory” if transfusion is deemed mandatory to succeed in the optimum Hb degree primarily based on affected person information from time level t to t + 3. Conversely, it outputs “non-necessary” if transfusion is deemed pointless.

Experimental setup

Following the outline supplied within the latter a part of Fig. 1, the coaching, validation, and take a look at datasets have been partitioned into ratios of 0.7, 0.1, and 0.2, respectively, to make sure that there have been no overlapping patients throughout datasets. Subsequently, every variable was standardized by remodeling it to a imply of 0 and a typical deviation of 1. The variables have been reshaped into sequential configurations and aligned with the enter necessities of the proposed mannequin. The chosen window measurement for this transformation was established as 4, primarily based on the findings derived from the experimental outcomes offered in Appendix Desk 2. The ultimate construction of the enter information was organized as (data_num, window_size, feature_num).

To coach the proposed mannequin, we set the next parameters: 500 epochs, batch measurement of 128, studying price of 0.0005, and dropout price of 0.5. The chosen optimizer was Adam, and the loss perform utilized was categorical cross-entropy. Early stopping was applied throughout mannequin coaching to effectively handle the computing assets and cut back pointless useful resource utilization. The mannequin with one of the best efficiency was chosen for subsequent analyses.

Within the experiments outlined in Desk 1, the epoch, batch measurement, and studying price have been set to 500, 128, and 0.003, respectively. For XGBoost, the n_estimator was set to 500. Within the case of the MLP, the Adam optimizer was used, with a dropout price of 0.6. Adam was additionally adopted because the optimizer for the LSTM however with a dropout price of 0.5. All hyperparameters have been decided by parameter tuning to attain optimum efficiency.

To evaluate the predictive capabilities for Hb ranges, linear regression, XGBoost¹³, and MLP have been used as typical machine studying benchmarks. Moreover, to evaluate the capability to include temporal dependencies intrinsic to sequential information from patients undergoing persistent dialysis, we included LSTM, a variant of the RNN, and the GRU structure, that are the Hb prediction fashions beforehand proposed by Yun⁹. Particularly, we used the unique GRU mannequin denoted as “GRU structure (A)” and one other variant denoted as “GRU with GNR (Baseline).”

Statistical evaluation

The Hb degree prediction activity was thought of a regression drawback, and its efficiency analysis utilized the next 4 metrics: R-squared, MSE, RMSE, and MAE. Conversely, within the context of the ESA dose advice and transfusion alarm duties, that are thought of classification issues, the analysis metrics included precision, recall, F1-score, and accuracy. Experiments have been performed using Keras with a TensorFlow backend.