A predictive model for real-time LSTM methods: Monitoring dynamic transmembrane pressure improves loop life and anticoagulant therapy accuracy in continuous renal replacement therapy

Abstract

Background

Continuous Renal Replacement Therapy (CRRT), is essential for managing acute kidney injury (AKI) Dynamic monitoring of transmembrane pressure (TMP) during CRRT is crucial for predicting filter clotting and optimizing filter lifespan, which indirectly supports anticoagulation management.

Objective

To prolong the lifespan of CRRT circuits and enhance the precision of anticoagulation therapy by developing a predictive early warning model for CRRT circuit life, based on dynamic TMP monitoring.

Methods

We conducted a retrospective analysis in the ICU of the First Affiliated Hospital of Army Medical University. Leveraging the TMP data recorded by CRRT machines, we established an adaptive real-time predictive modeling framework, termed DTP (Dynamic Transmembrane Pressure Prediction), utilizing Long Short-Term Memory (LSTM) networks. This framework predicts TMP trends as an early indicator of filter clotting. Our models were validated using over 20,000 min of clinical data from 405 CRRT cases, predicting TMP trajectories within 50 min.

Resuits

In simulated treatment evaluations, our LSTM models accurately identified impending TMP increases, achieving recall rates exceeding 0.97 and F2 scores above 0.93. Notably, an average warning time of 23 min was provided prior to the TMP reaching the critical 260 mmHg threshold, indicating substantial filter clotting. An analysis of false alarms revealed patterns consistent with emerging instability and transient artifacts.

Conclusion

The personalized early warning model developed within the DTP framework effectively predicts TMP changes, enhancing the accuracy and timeliness of medical interventions. This improvement reduces the incidence of adverse events, maximizes the lifespan of CRRT circuits, and ultimately decreases treatment and personnel costs.

Keywords

predictive model long short-term memory (LSTM) algorithm continuous renal replacement therapy transmembrane pressure (TMP)dynamic transmembrane pressure monitoring adverse event prevention

Introduction

Acute kidney injury (AKI) is an independent risk factor for both incidence and mortality, with an incidence rate in the intensive care unit (ICU) that can be as high as 50%^.^1,2 Continuous Renal Replacement Therapy (CRRT) is a blood purification technique commonly used in critically ill patients with AKI to provide continuous hemodynamic support and solute clearance over an extended period of time.³Current guidelines prioritize continuous therapy over intermittent therapy in patients with hemodynamic instability.^3–6 Up to 10–20% of critically ill patients receive CRRT during ICU admission, with clinical conditions necessitating CRRT, including sepsis, heart failure, liver failure, and postoperative fluid overload.^7,8 In critically ill patients undergoing CRRT, the use of heparin for systemic anticoagulation or citrate for regional anticoagulation is required to prevent circuit and filter clotting. However, determining the optimal anticoagulation therapy for critically ill patients poses a challenge, as excessive anticoagulation increases the risk of bleeding complications, whereas inadequate anticoagulation can lead to early circuit clotting, with an incidence rate ranging from 67.6% to 74.6%.⁹ A significant complication of CRRT is premature coagulation and clotting of the extracorporeal circuit, particularly within the hemofilter, where hemofilter membrane clotting is associated with an increase in transmembrane pressure (TMP).¹⁰ Deep learning techniques have increasingly become the cornerstone for developing predictive models that enhance clinical decision-making. Models that capture sequential data, such as recurrent neural networks (RNNs), are particularly effective for conditions such as Parkinson's, Alzheimer's, and heart disease^11–13 because of their proficiency in integrating the temporal dynamics of disease progression. Among these, sophisticated RNN variants such as Long Short-Term Memory (LSTM) networks have demonstrated superior performance in forecasting outcomes of opioid-related illnesses.^14–16

A substantial body of research indicates that LSTM networks have achieved state-of-the-art performance in applications ranging from natural language processing to time-series forecasting. The introduction of LSTM networks in the field of CRRT represents a paradigm shift from traditional threshold-based monitoring to predictive data-driven approaches. Our study leverages LSTM's capability to handle time-series data and offers a novel approach for predicting trends in TMP. To date, the potential of LSTMs to forecast CRRT circuit pressure trends for predicting impending filter clotting has not been explored.

We hypothesize that compared to simple threshold alarms, the LSTM approach can accurately model TMP trajectories and provide early warnings for increased circuit clotting risk. This study tested this hypothesis by assessing model performance metrics for TMP prediction using accurate clinical CRRT data.

Materials and methods

Dataset description

This retrospective observational study was conducted at the Department of Critical Care Medicine of the First Affiliated Hospital of the Army Military Medical University in China. Medical records and CRRT data were collected from January 1, 2019, to March 1, 2022. This study was approved by the Ethics Committee of the First Affiliated Hospital of the Army Military Medical University in China (BKY202250). A total of 405 CRRT episodes were included in the 298 patients. These episodes represented 25,132.63 h of treatment (Figure 1), Among them, 121 cases were unplanned disembarkation, with an incidence of 40%. Inclusion criteria were patients aged > 18 years, receiving CRRT treatment during hospitalization and unplanned discontinuation, complete records of routine blood tests, coagulation function, blood gas analysis, liver and kidney function, and complete records of CRRT data during treatment. Exclusion criteria were as follows: automatic discharge and disembarkation during hospitalization, death during hospitalization resulting in disembarkation, disembarking from the hospital, and use of other blood purification therapies, such as plasma exchange. Patients were treated with double vena cava catheters for vascular access, and all enrolled patients were treated with sodium citrate anticoagulation and continuous venovenous hemofiltration (CVVH) in pre-dilution mode. CVVH was performed using the Prismaflex System (Gambro) machine equipped with several types of filters, including ST100, M150, All femoral venous access was achieved through a 13 Fr double-lumen catheter (Baxter, USA), with blood flow maintained at 150–200 mL/min, the main components of the replacement solution(sodium 140 mmol/L, potassium 2.0 mmol/L, calcium 1.5 mmol/L, magnesium 0.75 mmol/L, chloride 110 mmol/L, and bicarbonate 32 mmol/L),replacement fluid rate set at approximately 30 mL/kg/h,Calcium ions are maintained at 0.25–0.45 mmol/L after filtration. and ultrafiltration rate determined by a clinical physician based on the patient's condition. No serious life-threatening complications occurred during CRRT treatment.

Figure 1.

Inclusion and exclusion criteria for study subjects.

Data pre-processing

The research data were collected at intervals of one min from the Prismaflex CRRT machine (Gambro, United States). The pressure variables included the minute AOP, EP, PFP, and RIP from the relevant circuit points.TMP, corresponding to the pressure of the filter membrane, was calculated from these data using the equation TMP = (PFP + RIP)/2–EP. To exclude outliers, the data were preprocessed using the Kalman filter algorithm.¹⁷To provide sufficient data for model training, 600 min of TMP values were generated after startup as the pre-training set for parameter tuning and as part of the continuously updated training sets later. The later predictions used sliding-window processing, and the size of the sliding window was 50 min. Each data sequence consisted of consecutive TMP values recorded every 50 min. The Figure 2 illustrates the results of TMP data processed through Kalman filtering, where the red line represents the processed data and the grey line represents the raw data.

Figure 2.

TMP data pre-processing.

In this study, we selected the Kalman filter algorithm to preprocess TMP data due to its dynamic estimation capability and robustness against noise. TMP data in the CRRT process are subject to fluctuations caused by patient physiological variations and measurement errors, which may introduce noise and outliers. The Kalman filter, by combining historical data with current observations, effectively removes noise while preserving the dynamic trends of TMP, ensuring the smoothness and reliability of the data. Additionally, the Kalman filter has low computational complexity, making it suitable for real-time data processing during CRRT.

During the research process, we also considered other data preprocessing methods, such as moving average filtering, median filtering, and low-pass filtering. However, these methods exhibited certain limitations when applied to TMP data. For instance, moving average filtering may overly smooth dynamic trends, median filtering is less effective for continuous noise, and low-pass filtering may introduce lag effects in capturing TMP dynamics. In contrast, the Kalman filter strikes a better balance between noise removal and trend preservation. Experimental results further validated its effectiveness, showing that TMP data preprocessed with the Kalman filter significantly improved the predictive performance of the LSTM model (e.g., F2-Score and AUC), confirming its suitability for this study.

In our study, TMP values were recorded every 50 min. This frequency was determined as part of the standard institutional protocol for CRRT monitoring at our facility. The 50 min interval reflects a balance between capturing sufficient temporal resolution to monitor dynamic changes in TMP and minimizing unnecessary data redundancy or excessive workload for clinical staff. This protocol ensures consistent and reliable data collection during CRRT treatment and aligns with routine clinical practices.

While the recording frequency was not specifically designed for this study, it proved to be suitable for our predictive modeling approach. The temporal resolution provided by the 50 min interval allowed the LSTM model to effectively capture the temporal dynamics and long-term trends in TMP changes, which are critical for predicting filter clotting risk. Future studies may explore the impact of different recording intervals on model performance to further optimize data collection strategies.

Methodology

In this study, the high heterogeneity of the patient population posed a significant challenge. This heterogeneity is characterized by diverse disease types and substantial physiological variations among individuals, making the quantification and modeling of coagulation factors a complex task. Traditional single models often struggle to adapt to such heterogeneity, and may fail to capture the unique physiological characteristics of each patient and their impact on coagulation. Therefore, a unique approach was employed, in which an individual model was established for each patient receiving CRRT treatment. This approach utilizes the similar neural network structure for all models; however, each model independently learns and adjusts its parameters to best fit the specific data of the respective patient. The advantage of this strategy is its ability to capture and learn subtle differences that represent individual physiological states. As a result, it provides more accurate circuit clotting warnings tailored to each patient. The study outlines the model's pretraining, parameter optimization, and online learning processes in detail, showcasing how these components are seamlessly integrated into an automated modeling workflow. This approach allows the construction of a modeling framework called DTP that combines individualization with process consistency.

Model construction

First, the model chosen by the Modeling Framework is LSTM, LSTM networks (Figure 3) are a particular class of RNN having special memory cells working as an information accumulator together with a circuit of input, output, and forget gates^.^18,19 These networks can learn both short and long-term dependencies and make a prediction based on time series data when there is a random lag between two consecutive events.

Figure 3.

The schematic of Long Short-Term Memory.

The model architecture consists of multiple layers of LSTM units, with each LSTM layer containing a set of LSTM nodes. The raw TMP dataset, after being divided into sequences composed of 49 consecutive TMP values, forms individual input data sequences. Let n be the layer number and m be the node number in each layer, each of these sequences passes through each LSTM layer, resulting in the generation of a [49*m] matrix at each layer. This matrix serves as the output for the current layer and the input for the next layer. The final layer's output is further processed through a fully connected layer to produce a single predicted TMP value. This predicted value is then compared to the actual TMP value using the Mean Squared Error (MSE) loss function to assess the prediction accuracy. Parameter adjustments are made through backpropagation with a learning rate of v. The stacking of multiple LSTM layers leads to a significant increase in the number of model parameters as the LSTM layers increase in number. This poses a risk of overfitting and prolonged training times. To mitigate these issues, the DTP framework incorporates dropout layers between each LSTM layer to prevent overfitting, with the dropout rate set to k.

Pre-training and optimization

Therefore, while the structural architecture remains consistent across models, the values of n, m, v, and k serve as hyperparameters that can vary and are determined by pre-training. To achieve the best prediction performance, the hyperparameters including dropout, learning rate, the number of neurons in the LSTM layer, and the number of LSTM layers were optimized. Considering the complexity of parameter combinations, the OPTUNA optimization framework²⁰ was used to optimize a tree-based hyper-parameter based on the Tree-structured Tarzen Estimator (TPE) sampler^,²¹ which relied on Bayesian optimization^21,22 to determine the most promising hyperparameter combination after one hundred iterations. For precise prediction, the initial 600 TMP values of each patient were used for pre-training, hyperparameters optimization, and individual model construction.

Rolling forecast and online learning

Once each model has determined its hyperparameter combination through pre-training, its structural architecture remains unchanged and the prediction for TMP officially launchs. The prediction mechanism here applied the rolling forecast mechanism,^23–25 which is shown in Figure 4 above, and involves predicting the TMP for the next 50 min. It starts by using the TMP values from the previous 49 min to predict the TMP for the next minute. Then, it combines the predicted TMP value with the TMP values from the preceding 48 min and continues this iteration for a total of 50 steps. Therefore, the models trained within this framework have a prediction cycle of 50 min. During one prediction cycle, the model's parameters remain fixed.

Figure 4.

The rolling forecast mechanism. The model computes the first predicted value based on an input consisting of the most recent 49 true values, then places the predicted value after these true values and removes the first true value to obtain a new input, the new input is used to compute the second predicted value, and so on, finally producing 50 predictions.

However, the model's parameters are updated through online learning at fixed time intervals. After completing a prediction cycle, to promptly adapt to TMP variations, the DTP framework automatically guides the model into the online learning phase. The process involves introducing the actual TMP values from the past 50 min as new data into the original training dataset. With the model's structural architecture (i.e, hyperparameter combination) unchanged, it undergoes training on the updated training dataset and proceeds into a new prediction cycle. The optimal weight matrix is selected with the early stopping method.^26–28

At each moment within a prediction cycle, the model compares the predicted TMP value with an alarm threshold. If the predicted value exceeds the threshold, an alarm is triggered. Since the prediction cycle spans 50 min, an alarm signal indicates the TMP change expected in 50 min, thus it is considered valid only if the actual TMP reaches the alarm threshold within 50 min after the alarm is triggered. The detailed algorithm flow is shown in Figure 5.

Figure 5.

The TMP monitoring and predicting flow chart.

Experiment

Experiment setting

Our DTP framework was utilized to model and test all 405 cases, revealing that as long as patients could safely disengage from the machine without any TMP anomalies, our predictive model did not trigger an alert. To objectively evaluate the quality of the model developed through the framework of automatic modeling, we focused on analyzing the predictive performance of the model on the cases with elevated TMP. Our implementation environment was the Python programming language (3.8). Filtering algorithms were implemented with the Python PyKalman package. Deep learning was implemented with Python TensorFlow,²⁹ Python Keras³⁰ and PyTorch Geometric^.³¹ Other libraries used included Python NumPy,³² Python Pandas,³³ and Python SHAP^.³⁴

Prediction evaluation

In clinical practice, a TMP reaching or exceeding 200 mmHg is typically considered a warning threshold for the potential risk of clotting. However, this threshold may vary depending on specific clinical scenarios and differences in medical equipment. Therefore, in the evaluation of the model's performance, we selected warning thresholds for the model ranging from 200 mmHg to 300 mmHg, with intervals of 5 mmHg, to comprehensively assess the model's alerting effectiveness.

For each model in which the DTP framework was automatically modeled, Precision, Recall rate, and F-score^35,36 were employed to reflect the quality of the predictions. Precision is the proportion of all positive predictions that were correctly predicted. The Recall rate is the proportion of correct predictions that are positive for all actual positives. When the filter is running, if the TMP reaches or exceeds the specified threshold after a prediction cycle (50 min), there is a hidden danger and it is running abnormally, and the model should send an alarm signal before the failure of anticoagulation and clogging occurs. At this moment, if the model provides an alarm, then it is defined as a True Positive (TP). Otherwise, missed alarm detection is recorded as a False Negative (FN). On the other hand, if TMP fluctuates smoothly between 0 mmHg and the threshold within a forecast period, meaning that the machine is in good working order, an alarm signal given in such a case is defined as a False Positive (FP). When the filter is in good condition and the model is silent without any alarm, the situation is True Negative (TN). Multiple alarm signals without intervals are recorded as one alarm to mitigate alarm fatigue among healthcare professionals, streamline clinical workflows, and maintain system performance. This approach ensures that critical alerts are given due attention, reduces unnecessary interruptions, and enhances patient safety by minimizing stress and disturbances. It also aids in data analysis to better understand TMP trends and improve the overall cost-effectiveness of healthcare resource utilization. Because the Precision and the Recall rates affect each other^,²⁵ the pursuit of a high precision rate results in a low recall rate, and vice versa. F-score was used to comprehensively evaluate them. The calculation formula is:

F_{β} = (1 + β^{2}) \times \frac{P r e c i s i o n \times R e c a l l}{β^{2} \times P r e c i s i o n + R e c a l l}

Here, β is a parameter that balances Precision and Recall. When β = 1, both metrics are equally important, resulting in an F1-score. For our current situation, where minimizing missed detections (False Negatives, FN) is crucial, we prioritize recall. Therefore, we opted for the F2-score with β set to two, emphasizing the significance of the recall rate. To comprehensively assess the DTP framework's efficacy, we consolidate the counts of True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN) from all deployed models. Leveraging this collated data, we meticulously calculate key performance metrics: Recall, Precision, and the F2-score. Additionally, for a nuanced visual representation of the framework's alerting proficiency across diverse thresholds, we meticulously plot a Receiver Operating Characteristic (ROC) curve. This analytical strategy furnishes a holistic and nuanced insight into the predictive accuracy and overall robustness of the framework.

Advanced time

After the model starts, when the actual TMP value exceeds the threshold, the CBP enters an abnormal state corresponding to the current threshold.^37,38 The latest alarm is considered to have the strongest correlation with this unusual state. If the time interval between them was less than 50 min, the alarm was judged to be effective, and this time interval was recorded as the early alarm advance time. To assess the efficacy of the DTP framework's warnings under different threshold settings, we calculated the mean of all effective early alarm advanced times across all the models. Given that the prediction cycle is set at 50 min, a mean early alarm advanced time closer to 50 min signifies a higher effectiveness.

Failure analysis

Detailed statistical analysis was conducted to assess the occurrence rates of False Negatives (FN) and False Positives (FP). FN events represent instances where an alarm failed to detect an actual event, whereas FP events denote instances where the alarm signaled an event that did not occur. We aimed to provide a comprehensive understanding of the performance and reliability of the framework by systematically recording FN and FP occurrences across various operational scenarios and conditions. Through careful categorization and analysis of FN and FP occurrences, we identified patterns and trends associated with different operational contexts, enabling a nuanced evaluation of system failures.

Results

In our study, 298 patients were screened for 405 episodes of CRRT, and the effective treatment time was 25,132.63 h. Among them, 258 patients with TMP < 250 mmHg had 360 episodes of CRRT, all of which were scheduled to be discontinued. Forty patients with TMP > 250 mm Hg had 45 CRRT episodes, all of which were unplanned disembarkations. The demographic characteristics of the 40 selected patients with TMP > 250 mm Hg are shown in Table 1, and the experimental data of the 40 patients before CRRT onset are shown in Table 2.

Table 1.

Displays the demographic and treatment characteristics of the selected 40 patients.

	Patients (n = 40)
Gender (male/female)	26/14
Age	52.82 ± 19.53
Height	164.9 ± 7.43
weight	66.68 ± 10.64
APACHE II	27.23 ± 8.66
Diagnosis
Respiratory disease	11
Cardiovascular disease	5
Digestive diseases	8
Sepsis	4
Diseases of urinary system	8
Autoimmune disease	4

Table 2.

Laboratory data of 40 patients before CRRT.

	Patients (n = 40)
HCT	28.23 ± 8.50
Platelet (10 $\land$ 9/L)	154.62 ± 99.98
APTT,s	42.53 ± 26.55
Fib(g/L)	4.24 ± 4.83
PT,s	18.48 ± 22.47
D-Dimer(mg/L)	11.06 ± 19.20
FDP(ug/ml)	19.15 ± 20.46
Uric Acid(mmol/L)	319.10 ± 188.62
ALP(U/L)	249.94 ± 792.84
Urea nitrogen (mmol/L)	17.69 ± 14.00
hemameba(mmol/L)	11.92 ± 6.42
ALT (μmol/L)	128.86 ± 268.45
AST (μmol/L)	158.61 ± 403.25
TBIL(μmol/L)	41.16 ± 93.86
IBIL (μmol/L)	18.5 ± 40.68
DBIL(μmol/L)	17.21 ± 47.66

Abbreviations: APACHE II, acute physiology and chronic health evaluation II; HCT,hemato cri tvalue;APTT, activated partial thromboplastin time; Fib, fibrinogen; PT, prothrombin time; FDP, fibrinofen degradation product; ALP, alkaline phosphatase; ALT,alanine aminotransferase;AST,aspartate aminotransferase;TBIL,total bilirubin;IBIL,indirect bilirubin;DBIL,direct bilirubin

Basic information of patients and disease distribution: As can be seen from Table 1, among the 40 patients, 26 were males and 14 were females, with certain differences in gender distribution. In terms of age, the average age was 52.82 ± 19.53 years old, which was in the middle age range as a whole, but the individual differences were large. In the diagnosis of diseases, covering a variety of types, respiratory diseases accounted for a relatively high proportion of patients, there are 11 cases; Followed by digestive system diseases and urinary system diseases, 8 cases respectively; The number of patients with cardiovascular diseases, sepsis and autoimmune diseases is relatively small. Table 2 shows that there are significant fluctuations in various blood indexes and biochemical indexes of the patient. Among the indicators related to coagulation function, such as APTT (42.53 ± 26.55 s), PT (18.48 ± 22.47 s)etc. The abnormalities of liver and kidney function indexes, such as ALT (128.86 ± 268.45 μmol/L), AST (158.61 ± 403.25 μmol/L) and urea nitrogen (17.69 ± 14.00 mmol/L).

Prediction result

To determine the optimal alarm threshold, we conducted tests using thresholds ranging from 200 to 300, at intervals of 5. We then computed and analyzed the evaluation metrics, as shown in Table 3. From the evaluation results obtained with different thresholds, the difference in threshold selection is prominently reflected in Precision and F2Score, whereas the performance gap in Recall is relatively small. Initially, when the threshold was set above 200 mmHg, recall consistently remained above 0.95, indicating that our framework's adaptive model effectively anticipated instances of high TMP, promptly issuing alerts for clinical intervention. However, significant differences in precision were observed across different thresholds. At a threshold of 250 mmHg, precision peaked at 0.8, whereas at thresholds of 205 mmHg and 220 mmHg, precision reached a nadir of 0.598. Overall, both excessively high and low thresholds can increase the occurrence of False Positives (FP), significantly impacting precision and consequently affecting the final F2Score. The optimal F2Score was achieved at a threshold of 260 mmHg, reaching 0.933. Clinically, TMP > 250 mmHg is a typical manifestation of circuit coagulation, indicating that our framework's modeling effectively predicts imminent circuit coagulation. To further analyze and determine the final threshold, we calculated the True Positive Rate (TPR) and False Positive Rate (FPR) for each threshold, and plotted the ROC curve accordingly (Figure 6). The ROC curve with an AUC of 0.92 indicates a reasonably good predictive performance.

Figure 6.

ROC curve with multiple thresholds for prediction evaluation.

Table 3.

Prediction performance comparison at different TMP thresholds.

TMP threshold(mmHg)	TP	FN	FP	Recall	Precision	F2Score
200	232	6	132	0.974	0.637	0.881
205	273	6	183	0.978	0.598	0.868
210	273	6	173	0.978	0.612	0.873
215	273	6	163	0.978	0.626	0.879
220	299	6	201	0.980	0.598	0.869
225	307	6	194	0.980	0.612	0.875
230	343	6	192	0.982	0.641	0.888
235	377	6	174	0.984	0.684	0.904
240	397	8	154	0.980	0.720	0.914
245	471	11	154	0.977	0.753	0.922
250	477	14	117	0.971	0.803	0.932
255	496	16	139	0.968	0.781	0.924
260	464	12	117	0.974	0.798	0.933
265	375	14	104	0.964	0.782	0.921
270	394	12	128	0.970	0.754	0.917
275	351	11	125	0.969	0.737	0.912
280	322	6	106	0.981	0.752	0.925
285	307	4	96	0.987	0.761	0.931
290	273	8	95	0.971	0.741	0.914
295	257	6	110	0.977	0.700	0.905
300	234	7	117	0.970	0.666	0.889

Advanced time

In analyzing the results of the model, we considered the temporal correlation and continuity of the abnormal TMP elevation events. If the time interval between two abnormal events is very short, this may indicate a correlation between these events or continuous anomalies caused by the same underlying reason. Therefore, when counting the number of abnormal events, consecutive abnormal events are usually considered as a single anomaly rather than counting each event separately. Specifically, we adopt the strategy of an abnormal event window. In this strategy, we define a time window within which all abnormal events are treated as a single anomaly. The size of this time window can be determined based on the specific application scenarios and requirements. It can be a fixed-length period or dynamically adjusted based on the characteristics and trends of data. In this study, to maintain consistency between adjacent windows and avoid double-counting the same TMP anomaly, we selected a fixed time window of 50 min. During the specific statistical analysis, other abnormal TMP events within 50 min after one TMP anomaly were ignored. Therefore, our advanced time analysis was based on partial data sampling.

In addition, the advance time between the effective alarm and the actual threshold attack was analyzed. It is more meaningful if the advance time is over 5 min because it offers clinicians more time to respond properly. Figure 7 shows the distribution of 191 meaningful advance times. Overall, the distribution of the advance time was relatively uniform. After calculation, the average advance time was 23 min.

Figure 7.

The distribution of the effective advanced time.

False alarm

To comprehensively evaluate the performance of our DTP framework, we plotted prediction curves for all cases and invited medical professionals to conduct a detailed analysis and classification of the generated alert signals. Our analysis focused on three main scenarios: ideal cases, alert failures, and general cases. In the ideal cases, the model successfully issued an alert in advance without any false positives or missed alerts. Alert failures refer to instances where the model failed to issue an alert signal when the actual TMP values exceeded the threshold. General cases covered instances where the model correctly issued alerts in advance, though in some cases, TMP remained within the safe range in the subsequent 50 min, leading to false alarms. We will discuss these scenarios in detail and further analyze the causes of false alarms.

In an ideal prediction example, as shown in Figure 8(a), the actual TMP value reached the 300 mmHg threshold after 1500 min from the start of prediction, and our model successfully issued an alert at minute 1500 without any false alarms observed. Figure 8(f) presents an example of an alert failure: in the purple area around 1100 min, the actual TMP value exceeded 300 mmHg, but the model did not issue an alert signal at this time. This indicates that the model predicted the TMP value would drop back to the safe range within 50 min, which did not match the actual situation. Such missed alerts are very rare. Figures 8(b) to (e) show more common scenarios: the model issued correct alerts in advance before TMP eventually reached the threshold, but TMP remained within the safe range in the subsequent 50 min in some cases of false alarms. Figure 9 selects and presents three typical cases of false alarms. In these cases, although the actual TMP values remained below the threshold within the 50 min window, our model issued false alarms.

Figure 8.

Predictions in some cases. The black line represents the actual TMP after Kalman filtering; the red line represents the predicted TMP calculated by the model; the blue line is the warning line, i.e., when predicting that the TMP will exceed the threshold (300mmHg), it will appear 300 mmHg, otherwise it will keep at zero.

Figure 9.

Three types of false alarms.

The typical false alarm events are described as follows. Type I: False alarms caused by sharp fluctuations in data over a short period. Although we had applied Kalman filtering to the raw data to reduce the impact of extreme values and outliers on the model, sharp fluctuations could still occur in the short term, leading the model to predict that TMP would rapidly exceed the threshold. However, TMP actually remained at a relatively safe level. Figure 9(a) shows a classic case of such a false alarm. Type II: False alarms caused by data approaching the threshold. This mainly occurred after the CRRT had been running for a longer time. TMP slowly increased over time and fluctuated around 200 mmHg, but did not rise to the 300 mmHg level. In these cases, our model might predict that TMP would soon increase to the dangerous threshold of 300 mmHg. However, in many instances, TMP could stably operate between 200–300 mmHg for a long period; during this time, all alerts were considered false alarms. This type of false alarm is demonstrated in Figure 9(b). Type III: False alarms when data were at low levels. TMP fluctuated within an undisputedly safe range, but the model issued a warning, categorizing such false alarms into a separate category, as shown in Figure 9(c).

The reason for the Type I false alarm is that the instrument is in an unstable operating state, which requires continuous attention. Type II false alarm is mainly caused by the fact that the filter is already overloaded and needs to be replaced. Therefore, if the TMP does not exceed 300 mmHg after a false alarm occurs, it can be classified in combination with the previous and subsequent data to obtain additional valuable information. The distribution of all false alarm types is also counted and Type I accounts for 10%, Type III for 11%, and Type II for 79%. It can be seen that most of the alarms belong to Type I and Type II, and the proportion of Type III alarms, which are entirely meaningless, is rather small.

Discussion

Clinical significance

Traditional simple logic alarm systems lack the ability to prevent impending circuit failure. The significance of our LSTM application lies in its ability to handle the unique temporal patterns of TMP data, thereby facilitating early intervention strategies to mitigate the risk of filter clotting. In contrast, our model's predictive capacity for filter clotting provides physicians with a valuable time window for intervention and treatment. On one hand, this may involve increasing the appropriate anticoagulant dosage or adjusting the type of anticoagulant; on the other hand, it allows time for flushing the blood filter to help remove accumulated blood clots and deposits on the filter surface, thereby maintaining filter permeability.

Elevated D-dimer /FDP (Table 2) and sepsis (4/40 patients) may have contributed to circuit coagulation.CRRT prescription: high displacement fluid rate and filter type (ST100/M150) may aggravate the increase of TMP. Renal insufficiency (increased urea/creatinine ratio) indirectly affects coagulation through uremic platelet dysfunction but is less important than anticoagulant efficacy.During the treatment of CRRT, there is a close correlation between TMP and filter coagulation. Studies have shown that for every 1 mmHg increase in TMP, the risk of clotting by the filter increases significantly.^39–41 Once TMP exceeds 250 mmHg, it usually means that significant clotting has occurred in the filter, at which point intervention must be performed, otherwise it will lead to early termination of CRRT therapy.^42,43Frequent filter changes can disrupt treatment continuity, reduce treatment effectiveness, significantly increase medical costs and medical staff workload, cause additional suffering to patients, and increase the risk of infection. Therefore, in the treatment of CRRT, accurate prediction of filter coagulation and timely intervention are of great significance for maintaining the stability of the filter, ensuring the treatment effect, reducing medical costs, and improving the prognosis of patients.

To prevent clotting, CRRT is often treated with anticoagulants, but determining the best anticoagulant regimen can be challenging for severely ill patients with complex conditions.^3,6 Excessive anticoagulation will increase the risk of bleeding, and insufficient anticoagulation will lead to premature clotting of the pipeline.⁵ Therefore, real-time monitoring of the coagulation status of cardiopulmonary bypass and adjusting the use of anticoagulants accordingly are crucial to guiding the optimal timing of anticoagulation.

The proposed DTP framework offers a solution to comprehensively address the clinical need for predicting circuit coagulation. First, by collecting and processing data at the minute level rather than the daily level, our framework achieves heightened temporal resolution and precision in predictions and monitoring. This fine-grained approach enables better capture of temporal fluctuations and variations, thereby enhancing sensitivity to events and trends. Secondly, individual differences among patients, including disease status, physiological characteristics, and treatment regimens, significantly influence circuit coagulation risk. Establishing a universal model to accommodate these variances was not feasible. However, our DTP framework allows for customization based on each patient's unique characteristics, thereby better addressing individual needs. The long-term benefit of our approach lies in the continuous monitoring and adaptation of the model. Fine-grained methods facilitate ongoing improvements in predictive performance by adapting to evolving patient conditions. This contributes to extending circuit lifespans, reducing the need for emergency replacements, and ultimately, lowering treatment costs.

The DTP framework predicts TMP changes in real-time through the LSTM model, triggering an alert on average 23 min before TMP reaches 260 mmHg (recall rate 94%), providing a critical time window for clinical intervention. Therefore, health care providers can adjust anticoagulation strategies (such as increasing the dose or changing the anticoagulant agent) or flushing the filter early to reduce the risk of clotting, prolong the use of the filter, and reduce unplanned discharge from the machine. This model has no false alarm when the patient's TMP is normal, and it has high prediction sensitivity for abnormal TMP, which significantly improves the continuity and safety of CRRT treatment.

The validity of the model is certain, but the results predicted by different CRRT model frameworks may be different, which needs to be further verified by the next trial.The CVVH pre-dilution mode used in this study, with a replacement fluid rate of 30 mL/kg/h, and the current model trained on high FF data, may be sensitive to high FF scenarios. If FF is reduced (such as the CVVHD mode), it may be necessary to further validate the ability of the model to capture TMP trends. The current DTP framework performs well in the CVVH mode with high FF and citrate anticoagulation (recall rate 94%), but further research may be needed for different CRRT prescription and patient populations in the future: The TMP warning value was adjusted according to the anticoagulation effect and hypercoagulable state. The sub-model was established according to CVVH/CVVHD/CVVHDF, and the real-time calibration was individualized.

The DTP framework has proven effective in generating TMP prediction models without triggering false alarms when patients can safely offline and TMP remains stable. Using the TPE (Tree-structured Parzen Estimator) method for automated network optimization significantly improves model performance. By efficiently tuning hyperparameters, TPE builds personalized LSTM models tailored to individual TMP trends, enhancing prediction accuracy and adaptability across different patients. Its dynamic adjustment capabilities also accelerate the optimization process and improve generalization.

Limitations and future work

One issue with the current DTP framework is that the predictive models generated through automatic modeling tend to have a relatively high false-alarm rate during predictions. According to the performance of TMP when an alarm signal occurs, three types of false alarms are classified. The first type is caused by sudden data fluctuations in the short term, the second type is triggered by TMP fluctuations near thresholds, and the third type occurs when the TMP is low. These false alarms may lead healthcare providers to take unnecessary action for patients in a safe state, requiring additional effort and consuming unnecessary medical resources. Fortunately, as shown in the Results section, false alarms mainly concentrate on the first two types, where the triggers are relatively clear. Therefore, when these two types of false alarms occur, it is meaningful for healthcare providers to take action. One limitation of our study is the lack of comprehensive analysis and identification of the third type of false alarm, which may result in some consumption of medical resources. In future research, a more thorough analysis of the causes of the third type of false alarm could be considered by incorporating additional factors.

Another significant challenge is the variability in the model's performance among patients with unique clotting patterns. This variability indicates the need for further model refinement through integration of multimodal data sources. Future research should explore the inclusion of additional physiological parameters, patient demographics, and treatment-specific factors to enhance the accuracy and applicability of the model across different patient populations. Additionally, expanding the training dataset to encompass a broader range of TMP trends can improve the generalization ability of the model, potentially reducing the occurrence of false alarms, as well as the integration of the model with existing CRRT systems. This integration can dynamically adjust treatment parameters based on real-time data, optimizing treatment outcomes. The practical application of LSTM models within CRRT machines can transform current treatment practices by directly providing real-time predictive insights to clinical physicians. Furthermore, the application of LSTM models in CRRT paves the way for research in other critical care domains, where similar approaches can be applied to other extracorporeal therapies and monitoring systems, such as Extracorporeal Membrane Oxygenation (ECMO), circulatory systems, respiratory systems, and neurological assessments, where predictive monitoring can significantly impact patient treatment outcomes.

Conclusion

In summary, we presented a novel DTP framework that accurately simulates TMP trajectories in critically ill patients undergoing CRRT, providing early warnings of increased circuit clotting risk. When TMP reaches the threshold of 260 mmHg, physicians have an average warning time of 23 min before filter clotting occurs.This model offers a valuable time window for clinical intervention, thereby improving the effectiveness of CRRT, prolonging its usage, reducing complications, enhancing patient survival rates, and potentially lowering treatment costs in certain cases.

Footnotes

Acknowledgements

The authors wish to express their sincere gratitude to the patient and family for their time and co-operation.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was supported by The Chongqing Science-health joint medical research project (grant number 2022QNXM27) and The Senior Medical Talents Program of Chongqing for Young and Middle-aged.

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

The datasets used and analysed during the current study are available from the corresponding author on reasonable request.

References

Jiang

Zhu

Luo

, et al. Epidemiology of acute kidney injury in intensive care units in Beijing: the multi-center BAKIT study. BMC Nephrol 2019; 20: 468.

Meersch

Küllmar

Wempe

, et al. Regional citrate versus systemic heparin anticoagulation for continuous renal replacement therapy in critically ill patients with acute kidney injury (RICH) trial: study protocol for a multicentre, randomised controlled trial. BMJ Open 2019; 9: e024411–e024411.

Siebeck

Dimski

Brandenburger

, et al. Super high-flux continuous Venovenous hemodialysis using regional citrate anticoagulation: long-term stability of middle molecule clearance. Ther Apher Dial 2018; 22: 355–364.

Karkar

Ronco

. Prescription of CRRT: a pathway to optimize therapy. Ann Intensive Care 2020; 10: 32.

Küllmar

Zarbock

. Renal replacement therapy in acute kidney injury: from the indications to cessation. Anaesthesist 2019; 68: 485–496.

Tolwani

. Continuous renal-replacement therapy for acute kidN54ZWQ Zinjury. N Engl J Med 2012; 27: 2505–2514.

Saraiva

Ortiz-Soriano

Mei

, et al. Continuous renal replacement therapy in critically ill patients with acute on chronic liver failure and acute kidney injury: a retrospective cohort study. Clin Nephrol 2020; 93: 187–194.

Rewa

Villeneuve

Lachance

, et al. Qua1ity indi-cators of continuous renal replacement therapy (CRRT) care in critically ill patients: a systematic review. Intensive Care Med 2017; 43: 750–763.

Heise

Gries

Moerer

. Predicting restoration of kidney function of during CRRT-free intervals. J Cardiothorac Surg 2012; 7: 1–7.

10.

Michel

Ksouri

Schneider

. Continuous renal replacement therapy: understanding circuit hemodynamics to improve therapy adequacy. Curr Opin Crit Care 2018; 24: 455–462. 10.1097.

11.

Che

Xiao

Liang

, et al. An rnn architecture with dynamic temporal matching for personalized predictions of Parkinson's disease. In: Proceedings of the 2017 SIAM international conference on data mining. Society for Industrial and Applied Mathematics, 2017, pp.198–206.

12.

Pillai

Bee

Kiruthika

. Prediction of heart disease using rnn algorithm. Int Res J Eng Technol 2019; 4: 5–11.

13.

Cui

Liu

. Longitudinal analysis for Alzheimer's disease diagnosis using RNN. In: 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018), 2018, pp.1398–1401.

14.

Dong

Deng

Hou

, et al. Predicting opioid overdose risk of patients with opioid prescriptions using electronic health records based on temporal deep learning. J Biomed Inform 2021; 116: 103725.

15.

Dong

Rashidian

Wang

, et al. Machine learning based opioid overdose prediction using electronic health records. In: AMIA Annu Symp Proc, 2020, pp.389–398.

16.

Xinyu

Jianyuan

Sina

, et al. Identifying risk of opioid use disorder for patients taking opioid medications with deep learning. J Am Med Inform Assoc 2021; 28: 1683–1693.

17.

Kim

Fealy

Baldwin

, et al. Premature circuit clotting due to likely mechanical failure during continuous renal replacement therapy. Blood Purif 2010; 30: 79–83.

18.

Hochreiter

Schmidhuber

. Long short-term memory. Neural Comput 1997; 9: 1735–1780.

19.

Greff

Srivastava

Koutnik

, et al. LSTM: a search space odyssey. IEEE Trans Neural Netw Learn Syst 2017; 28: 2222–2232.

20.

Bishop

. Pattern recognition and machine learning. Secaucus, NJ: Springer, 2009.

21.

Bergstra

Bardenet

Bengio

, et al. Algorithms for hyper-parameter optimization. Adv Neural Inf Process Syst. 2011; 24: 102–108.

22.

Shahriari

, et al.

Taking the human out of the loop: a review of Bayesian optimization.

Proc IEEE. 2016; 104: 148–175.

23.

, et al. ENSO forecasting over multiple time horizons using ConvLSTM network and rolling mechanism. In: 2019 international joint conference on neural networks (IJCNN), 2019. IEEE.

24.

, et al. NAO index prediction using LSTM and ConvLSTM networks coupled with discrete wavelet transform. In: 2019 International Joint Conference on Neural Networks (IJCNN), 2019. IEEE.

25.

Yuan

, et al.

Typhoon intensity forecasting based on LSTM using the rolling forecast method.

Algorithms 2021; 14: 83.

26.

Prechelt

. Early stopping—but when? In: Neural networks: tricks of the trade: second edition. Proceedings of the IEEE. Berlin: Springer, 2012, pp.53–67.

27.

Caruana

Lawrence

Giles

. Overfitting in neural nets: Backpropagation, conjugate gradient, and early stopping. Adv Neural Inf Process Syst. 2000; 7: 34–37.

28.

Morgan

Bourlard

. Generalization and parameter estimation in feedforward nets: Some experiments. In: [22]. 1990, pp.630–637.

29.

Abadi

, et al. Tensorflow: A system for large-scale machine learning. In: 12th Symposium on operating systems design and implementation, 2016.

30.

Chollet

, et al. Keras. Available from: GitHub. 2015.

31.

Fey

Lenssen

JE.

Fast graph representation learning with PyTorch Geometric.arXiv preprint. arXiv:1903.02428, 2019.

32.

Bressert

. SciPy and NumPy: an overview for developers. USA: O'Reilly Media, Inc, 2012.

33.

McKinney

. pandas: a foundational Python library for data analysis and statistics. In: Python for high performance and scientific computing. Nat Commun. 2011, 14, 1–9.

34.

Lundberg Scott

Su-In

. A unified approach to interpreting model predictions. In: Proceedings of the 31st international conference on neural information processing systems (NIPS’17). Red Hook, NY, USA: Curran Associates Inc., 2017, pp.4768–4777.

35.

Goutte

Gaussier

. A probabilistic interpretation of precision, recall and F-score, with implication for evaluation. In: Advances in Information Retrieval: 27th European Conference on IR Research, ECIR 2005, Santiago de Compostela, Spain, March 21–23, 2005. Proceedings 27. Springer Berlin Heidelberg, 2005.

36.

Powers

DMW

. Evaluation: from precision, recall and F-measure to ROC, informedness, markedness and correlation. 2020.

37.

Ejaz

Komorski

Ellis

, et al. Extracorporeal circuit pressure profiles during continuous venovenous haemofiltration. Nurs Crit Care 2007; 12: 81–85.

38.

Zhang

Baldwin

Zhu

, et al. Automated electronic monitoring of circuit pressures during continuous renal replacement therapy: a technical report. Crit Care Resusc 2015; 17: 51–54.

39.

del Castillo

López-Herce

Cidoncha

, et al. Circuit life span in critically ill children on continuous renal replacement treatment: a prospective observational evaluation study. Crit Care 2008; 12: R93.

40.

Kakajiwala

Jemielita

, et al. Membrane pressures predict clotting of pediatric continuous renal replacement therapy circuits. Pediatr Nephrol 2017; 32: 1251–1261.

41.

Kakajiwala

Jemielita

John

, et al. Hughes4 membrane pressures predict clotting of pediatric continuous renal replacement therapy circuits. Pediatr Nephrol 2017; 32: 1251–1261.

42.

Baldwin

Fealy

. Nusing for renal replacement therapies in the intensive care unit:historical,educational and protocol review. Blood Purif. 2009; 27: 174–181.

43.

Saif

Sarhan

Elshennawy

. Deep-kidney: an effective deep learning framework for chronic kidney disease prediction. Health Inf Sci Syst 2024; 12: 3.