A machine learning approach for predicting treatment response of hyponatremia

Abstract

Hyponatremia leads to severe central nervous system disorders and requires immediate treatment in some cases. However, a rapid increase in serum sodium (s-Na) concentration could cause osmotic demyelination syndrome. To achieve a safety hyponatremia treatment, we develop a prediction model of s-Na concentration using a machine learning. Among the 341 and 47 patients admitted to two tertiary hospitals for hyponatremia treatment (s-Na <130 mEq/L), those who were admitted to the general unit with urine sodium <20 mEq/L or treated with desmopressin were excluded. Ultimately, 74 and 15 patients (342 and 146 6-hourly datasets) were included in the learning and validation data, respectively. We trained the prediction model using three regression algorithms for shallow machine learning to predict s-Na every 6 h during treatment with the data of patients with hyponatremia (median s-Na: 112.5 mEq/L; range: 110.0–116.8 mEq/L) from one hospital. The model was validated externally using the data of patients with hyponatremia (median s-Na: 117.0 mEq/L; range: 112.9–120.0 mEq/L) from another hospital. Using 5–7 predictors (water intake, sodium intake, potassium intake, urine volume, s-Na concentration, serum potassium concentration, serum chloride concentration), the support vector regression model showed the best performance overall (root mean square error = 0.05396; R² = 0.92), followed by the linear regression and regression tree models. The predicted s-Na levels, using explainable machine learning algorithms and clinically accessible parameters, correlated well with the actual levels. Thus, our model could be applied to the treatment of hyponatremia in clinical practice.

HYPONATREMIA, defined as a serum sodium (s-Na) concentration of <135 mEq/L, is the most common electrolyte disorder in clinical practice [1]. Hyponatremia occurs in 15%–30% of hospitalized patients [2-4], with higher frequencies observed in elderly individuals [3] and cancer patients [5]. Based on its extracellular fluid volumes, hyponatremia is classified into three subtypes: hypovolemic, euvolemic, and hypervolemic [6]. Hypovolemic hyponatremia is caused by vomiting, diarrhea, primary renal failure, and diuretic use, resulting in the depletion of extracellular fluid volumes. Euvolemic hyponatremia can be due to inappropriate antidiuretic hormone secretion and adrenal insufficiency secondary to hypopituitarism. Hypervolemic hyponatremia occurs with edematous diseases, including congestive heart failure, liver cirrhosis, and nephrotic syndrome. Although hyponatremia is classified into these three subtypes, the identification of the causes of hyponatremia, as well as differentiation between hypovolemic hyponatremia and euvolemic hyponatremia, remains challenging because of its complex pathology [7]. Therefore, the definitive diagnosis of etiology often remains elusive before treatment initiation, particularly during the early treatment phase [8].

Rapid s-Na correction with intravenous infusion of hypertonic saline is warranted when symptoms of hyponatremia due to cerebral edema are evident, because it could lead to cerebral herniation and even death [1]. However, rapid s-Na correction could cause osmotic demyelination syndrome (ODS), which manifests as impaired consciousness and tetraplegia. Prevention of ODS is crucial in the treatment of hyponatremia because ODS has no effective treatment and can be life-threatening [9]. To minimize the risk of ODS, a correction rate of s-Na of <8–10 mEq/L per 24-h period is recommended [10]. However, ODS may occur in patients with severe hyponatremia (frequency rate: 0.3%–1.1%) due to occasional rapid s-Na changes even when treated by a specialist [11].

In recent years, artificial intelligence technology has been increasingly applied in medical research to support diagnosis and predict treatment efficacy and prognosis in various fields. However, while studies have been conducted on hyponatremia to predict its onset in hospitalized patients and after pituitary surgery [12, 13], there has not been any research that reports the prediction of treatment-dependent s-Na concentration in patients with hyponatremia.

In the current study, to address the clinical need for the safe treatment of patients with severe hyponatremia, we trained a highly explainable shallow machine learning model that can predict the s-Na concentrations of patients undergoing fluid infusion treatment.

Materials and Methods

Patients

Patients admitted to two tertiary care hospitals, Nagoya University Hospital and the Japanese Red Cross Aichi Medical Center Nagoya Daini Hospital, for hyponatremia treatment between April 2015 and March 2020 were included in this study. Clinical data of patients treated in acute care units, where time series data can be obtained every few hours, were extracted from the electronic medical records of both hospitals for analysis. Patients with a urinary Na (u-Na) concentration of <20 mEq/L were excluded because their cause of hyponatremia was mainly dehydration and completely different from that of patients with a u-Na concentration of >20 mEq/L. Moreover, patients who were administered desmopressin (DDAVP) to prevent a rapid increase in s-Na during hyponatremia treatment [14] were excluded because their urine volumes could be affected. The machine learning model was trained and validated internally using data from Nagoya Daini Hospital as training data; it was validated externally using data from Nagoya University Hospital as test data.

Predictors used in the model

The following major laboratory values related to fluid volume and osmolality were extracted as candidates for predictors: s-Na, serum potassium (s-K), serum chloride (s-Cl), serum glucose, blood urea nitrogen, creatinine, estimated glomerular filtration rate, uric acid, total protein, albumin, hematocrit, plasma osmolality, u-Na concentration, urine osmolality, plasma renin activity, plasma arginine vasopressin, plasma aldosterone, serum cortisol, plasma adrenocorticotropic hormone, human atrial Na diuretic peptide, thyroid-stimulating hormone (TSH), free triiodothyronine (FT3), free thyroxine (FT4), urine volume, total water administered by infusion and drinking (water intake), Na administered by infusion (Na-IN), and potassium (K) administered by infusion (K-IN). Among these candidates, s-Na, s-K, s-Cl, water intake, Na-IN, K-IN, and urine volume were selected as predictors because they are frequently measured and recorded in practice when treating hyponatremia (Fig. 1A).

Fig. 1

Seven predictors and data processing methods.

(A) Seven predictors of the serum sodium (s-Na) concentration at 6 h (s-Na6h).

The s-Na6h was used as the response (T = t + 6). The predictors were s-Na, serum potassium (s-K), serum chloride (s-Cl), urine volume (water-OUT), water intake (water-IN), sodium infusion (Na-IN), and potassium infusion (K-IN). For the s-Na6h, only the measured values of s-Na were used; the complementary values were not used. The s-Na, s-K, and s-Cl values were those at the reference time (T = t). Water-OUT was the amount of urine output during the 6 h before the reference time (t – 6 to t). Water-IN, Na-IN, and K-IN were determined using the treatment information pertaining to the number of hours from the reference time (t to t + 6).

(B) Data processing methods.

All recorded times (in minutes and seconds) were recombined as 1-h units. To obtain laboratory values of the predictors, linear completion was performed to compensate for missing values. For water-IN, Na-IN, and K-IN, we referred to the infusion master data to determine the amounts of water, Na, and K administered each hour. For water-OUT, if the urine volume was not recorded hourly, then the urine volume was divided by the elapsed time from the previous recording time to distribute the urine volume. For the response (s-Na6h), only the measured values of s-Na were used; the complementary values were not used.

Response

We used the s-Na concentration at 6 h (s-Na6h) as the response. When treating hyponatremia, physicians typically perform blood examinations every 2–3 h during the acute phase to determine the s-Na concentration. We selected 6 h as the time interval for simulation based on the assumption that even if there was a large discrepancy between the s-Na predicted by the model and the actual value, it would be clinically manageable in 6 h.

Data preparation and preprocessing

Datasets extracted from the electronic medical records were converted to timetable data; one time was associated with each row for each patient, and timestamps were resampled as hourly units after removing the time window from admission to discharge. Regarding laboratory values among the predictors, linear interpolation was performed to compensate for the missing values to process the data for machine learning because the measurement intervals varied depending on the case and elapsed time.

The semi-structured data, such as intravenous infusion, are heterogeneous and require standardization. For example, infusion formulations have various compositions that may have multiple different names, depending on the pharmaceutical company, and some formulations may be mixed before administration. Therefore, we created master data for each infusion formulation and combination of formulations so that the dosage (mL) could be converted to the actual dosages of water (mL), Na (mEq), and K (mEq). Regarding water intake, Na-IN, and K-IN, we referred to the infusion master data to determine the amount of water, Na, and K administered each hour. Regarding urine volume, if it was not recorded hourly, then the urine volume was divided equally by the time elapsed from the previous recording time to distribute the urine volume.

Regarding s-Na6h after the reference time, only the measured values of s-Na were used; the complementary values were not used. The s-Na, s-K, and s-Cl values comprised the reference time values (T = t). Urine volume was the amount of urine output during the 6 h before the reference time (t – 6 to t). Water intake, Na-IN, and K-IN were determined using the treatment information pertaining to the number of hours from the reference time (t to t + 6) (Fig. 1B). During preprocessing of the data, normalization was performed by transforming the data so that all measurements had values between 0 and 1.

Machine learning model

We trained and tested the following three regression algorithms for shallow machine learning: linear regression, decision tree, and support vector regression (SVR). All these models were implemented using MATLAB version R2022a, which is a commercially available mathematics software and programming language tool commonly used for machine learning (MathWorks, Inc., Natick, MA, USA). To predict s-Na6h, we trained the model using only the data derived from Nagoya Daini Hospital and performed 10-fold cross-validation. Then, we examined the differences in accuracy between the training models based on combinations of 5–7 predictors. Subsequently, the accuracy of the learning models was externally validated using the Nagoya University Hospital dataset. During this study, shallow machine learning was selected for modeling because the number of datasets was too small for deep learning, owing to its relatively high explanatory potential compared to deep learning, and because it could be used as a benchmark for comparisons of the accuracy of other models in the future.

Ethical approval

This research has been complied with all the relevant national regulations and institutional policies and according to the tenets of the Helsinki Declaration and has been approved by Nagoya University Hospital Institutional Review Board (approval number 2020-0384).

After extracting data from the electronic medical records, pseudonymization and storage were performed in accordance with the guideline associated with the Japanese law regarding the protection of personal information and the protocol approved by the ethics committee.

Results

Learning data

During the 5-year period, from April 2015 to March 2020, a total of 341 patients were admitted to the Department of Endocrinology and Diabetes of Nagoya Daini Hospital for hyponatremia treatment (Fig. 2A). Of these, 234 patients treated in the general unit were excluded. For patients treated in intensive care units (ICUs) and other acute care units using critical care progress charts, it was possible to retrospectively collect information regarding infusion therapy and fluid delivery. Ten patients with u-Na <20 mEq/L, suggesting dehydration, were excluded. The u-Na concentration of one patient was not measured. Moreover, 22 patients who received DDAVP to prevent a rapid increase in s-Na during hyponatremia treatment were excluded. Finally, 342 6-hourly datasets of 74 patients were used as learning data to train the model.

Fig. 2

Study flow diagram.

(A) Learning data. A total of 341 patients were admitted to the Department of Endocrinology and Diabetes of Nagoya Daini Hospital for hyponatremia treatment. We excluded 234 patients treated in the general unit. Ten patients with u-Na <20 mEq/L and one patient without u-Na measurement data were excluded. Additionally, 22 patients who received desmopressin (DDAVP) to prevent overcorrection of s-Na were excluded. Finally, 342 6-hourly datasets of 74 patients were used as learning data to train the model.

(B) Validation data. A total of 47 patients were admitted to the Department of Endocrinology and Diabetes of Nagoya University Hospital for hyponatremia treatment. Thirty-one patients treated in the general unit were excluded. One patient with u-Na <20 mEq/L was excluded. No patients were administered DDAVP during treatment. Finally, 146 6-hourly datasets of 15 patients were used as validation data for external validation of the model.

Validation data

To obtain validation data, target patients were selected using the same criteria as those used for patients at Nagoya Daini Hospital (Fig. 2B). During the 5-year period, from April 2015 to March 2020, 47 patients were admitted to the Department of Endocrinology and Diabetes of Nagoya University Hospital for hyponatremia treatment. Of these, 31 patients treated in the general unit were excluded. For patients treated in the emergency departments and ICUs using critical care progress charts (Fortec ACSYS; Koninklijke Philips N.V., Eindhoven, the Netherlands), it was possible to retrospectively collect information regarding infusion therapy and fluid delivery. Only one patient with u-Na <20 mEq/L, suggesting dehydration, was excluded. No patients were administered DDAVP during treatment. Finally, 146 6-hourly datasets of 15 patients were used as validation data to externally validate the model.

Patient characteristics

Table 1 summarizes the characteristics of patients included in the learning and validation datasets. Data are expressed as median (interquartile range [IQR]) or number (percentage). For patients included in the learning datasets, the median age was 78.0 years (IQR: 69.3–86.0 years), and 64.9% were women. On admission, the median s-Na concentration was 112.5 mEq/L (IQR: 110.0–116.8 mEq/L), median plasma osmolality was 238.5 mOsm/L (IQR: 228.3–245.0 mOsm/L), median u-Na concentration was 62.0 mEq/L (IQR: 43.3–94.8 mEq/L), and median urine osmolality was 362.0 mOsm/kg (IQR: 259.5–510.8 mOsm/kg).

Table 1

Clinical characteristics of the study population

	Learning data	n = 74	Validation data	n = 15	p-value
Sex, female	48 (64.9%)	74/74	10 (66.7%)	15/15	0.894
Age (years)	78.0 (69.3–86.0)	74/74	76.0 (73.0–78.0)	15/15	0.985
Hospitalization (days)	11.5 (9.0–16.8)	74/74	13.0 (9.0–15.0)	15/15	0.793
Na (mEq/L)	112.5 (110.0–116.8)	74/74	117.0 (112.0–120.0)	15/15	0.179
K (mEq/L)	3.90 (3.50–4.60)	74/74	4.10 (3.90–4.75)	15/15	0.490
Cl (mEq/L)	81.5 (77.0–85.0)	74/74	83.0 (79.0–86.5)	15/15	0.667
BUN (mg/dL)	12.35 (9.50–18.85)	74/74	11.10 (7.95–14.30)	15/15	0.008
CRN (mg/dL)	0.610 (0.470–0.908)	74/74	0.590 (0.480–0.700)	15/15	0.248
eGFR (mL/min/1.73 m2)	71.70 (49.13–99.80)	74/74	84.50 (66.60–108.95)	15/15	0.435
UA (mg/dL)	3.145 (2.190–5.393)	68/74	2.50 (1.70–3.60)	15/15	0.016
TP (g/dL)	7.13 (6.60–7.62)	74/74	7.00 (6.50–7.25)	15/15	0.447
ALB (g/dL)	3.99 (3.49–4.42)	72/74	3.90 (3.50–4.20)	15/15	0.829
HCT (%)	33.55 (30.33–35.65)	74/74	32.40 (30.75–35.65)	15/15	0.943
GLU (mg/dL)	125.0 (103.3–146.0)	74/74	105.0 (98.0–181.5)	15/15	0.757
p-Osm (mOsm/L)	238.5 (228.3–245.0)	74/74	245.8 (236.0–249.0)	15/15	0.197
F (μg/dL)	23.90 (17.00–35.00)	73/74	12.00 (8.00–20.85)	15/15	<0.001
ACTH (pg/mL)	26.70 (13.45–45.50)	71/74	22.30 (13.10–33.30)	15/15	0.007
PAC (pg/mL)	126.50 (77.70–199.45)	71/74	116.50(75.40–136.75)	12/15	0.351
PRA (ng/mL/h)	1.20 (0.40–2.05)	71/74	1.50 (0.80–2.78)	12/15	0.337
TSH (μIU/mL)	1.560 (0.840–2.690)	73/74	1.361 (0.782–2.342)	15/15	0.780
FT3 (pg/mL)	2.160 (1.730–2.640)	73/74	2.000 (1.800–2.270)	14/15	0.253
FT4 (ng/dL)	1.570 (1.400–1.700)	73/74	1.220 (0.965–1.325)	15/15	<0.001
AVP (pg/mL)	1.60 (0.90–4.20)	69/74	0.70 (0.60–2.03)	14/15	0.001
u-Na (mEq/L)	62.0 (43.3–94.8)	74/74	73.0 (46.5–95.0)	15/15	0.956
u-Osm (mOsm/kg)	362 .0 (259.5–510.8)	72/74	476.0 (360.0–505.5)	11/15	0.557

Data are expressed as median (interquartile range) or n (%). All statistical tests were two-sided, and significance was defined as a p-value of <0.05 (in bold).

ACTH, adrenocorticotropic hormone; ALB, albumin; AVP, arginine vasopressin; BMI, body mass index; BNP, brain natriuretic peptide; BUN, blood urea nitrogen; Cl, chloride; CRN, creatinine; eGFR, estimated glomerular rate; F, cortisol; FT3, free triiodothyronine; FT4, free thyroxine; GLU, glucose; hANP, human atrial sodium diuretic peptide; HCT, hematocrit; K, potassium; Na, sodium; PAC, aldosterone; PRA, renin activity; p-Osm, plasma osmolality; TP, total protein; TSH, thyroid-stimulating hormone; UA, uric acid; u-Na, urinary sodium; u-Osm, urine osmolality.

For patients included in the validation data, the median age was 76.0 years (IQR: 73.0–78.0 years), and 66.7% were women. On admission, the median s-Na concentration was 117.0 mEq/L (IQR: 112.0–120.0 mEq/L), median plasma osmolality was 245.8 mOsm/L (IQR: 236.0–249.0 mOsm/L), median u-Na concentration was 73.0 mEq/L (IQR: 46.5–95.0 mEq/L), and median urine osmolality was 476.0 (IQR: 360.0–505.5). Significant differences (p < 0.05; two-sample t-test) were observed between learning data and validation data for the following items: blood urea nitrogen, uric acid, cortisol, FT4, and arginine vasopressin (AVP). None of the patients included in both datasets had hyperglycemia (blood glucose concentrations >250 mg/dL). No patients presented with symptoms suggestive of ODS during or after treatment. None of the patients died. All patients were discharged home or to a rehabilitation facility.

Predictors

Fig. 3A shows the candidate predictors and training datasets during the period from ICU admission to discharge. The data window duration was set to 1 h. The frequency of each feature was calculated by dividing the total number of records by the total number of data windows for all patients, resulting in the following seven predictors with the highest frequency: s-Na, s-K, s-Cl, water intake, Na-IN, K-IN, and urine volume.

Fig. 3

Selection of predictors and distribution of normalized data.

(A) Number of records. The total number of training data windows (bar graph, right vertical axis) and the ratio of the number of training data records to the number of data windows of predictor candidates (line chart, left vertical axis) during the period from intensive care unit (ICU) admission to discharge are shown. The data window duration was set to 1 h.

(B) Normalized data distribution. This is a box-and-whisker diagram of the normalized data. The x indicates the mean, the middle horizontal line indicates the median, and the boxes indicate the first through third quartile ranges. The upper and lower ends of the vertical bars indicate the maximum and minimum values, respectively, excluding the outliers.

Data normalization

The seven selected predictors and one response were normalized as follows: water intake’ = water intake/1000; Na-IN’ = Na-IN/100; K-IN’ = K-IN/10; urine volume’ = urine volume e/1000; s-Na’ = (136 – s-Na)/30; s-K’ = (5.0 – s-K)/3; s-Cl’ = (98- s-Cl)/30; and s-Na6h’ = (136 - s-Na6h)/30. A distribution plot of the normalized data is shown in Fig. 3B. The normalized results were as follows: water intake, 98.54%; Na-IN, 97.95%; K-IN, 97.08%; urine volume, 98.83%; s-Na, 97.37%; s-K, 97.66%; s-Cl, 77.19%; and s-Na6h, 97.08%.

Internal verification

The results of performing 10-fold cross-validation with various combinations of 5–7 predictors are presented in Table 2. The learning model using linear regression resulted in a root mean square error (RMSE) of 0.036768 and coefficient of determination (R²) of 0.97 (Fig. 4A). The training model using SVR resulted in an RMSE of 0.037026 and R² of 0.97 (Fig. 4B). The training model using regression trees showed an RMSE of 0.056895 and R² of 0.93 (Fig. 4C). Among these three models, the linear regression model showed the highest accuracy, which corresponded to the root mean square deviation (RMSD) between the predicted and measured s-Na values was 1.11 mEq/L. The SVR model trained with five predictors (excluding s-K and s-Cl) showed the best accuracy among all combinations of predictors (RMSE of 0.036466 and R² of 0.97) (Table 2).

Table 2

Differences in the prediction accuracy of 10-fold cross-validation based on combinations of predictors

Predictors		Water-IN	●		●	●	●	●	●	●	●	●	●
		Na-IN	●	●		●	●	●	●	●	●		●
		K-IN	●	●	●		●	●	●	●			●
		Water-OUT	●	●	●	●		●	●	●	●	●	●
		s-Na	●	●	●	●	●		●	●	●	●	●
		s-K	●	●	●	●	●	●		●		●
		s-Cl	●	●	●	●	●	●	●		●	●
Modeling	Linear regression	RMSE	0.036768	0.037210	0.037191	0.036674	0.037295	0.11897	0.036604	0.036708	0.036519	0.036977	0.036500
		R2	0.97	0.97	0.97	0.97	0.97	0.69	0.97	0.97	0.97	0.97	0.97
	Support vector regression	RMSE	0.037026	0.037328	0.037171	0.036934	0.037388	0.11923	0.036789	0.036856	0.036624	0.037044	0.036466
	linear kernel c = 0.2337, γ = 0.0234	R2	0.97	0.97	0.97	0.97	0.97	0.69	0.97	0.97	0.97	0.91	0.97
	Regression tree model	RMSE	0.056895	0.066843	0.067919	0.067541	0.066918	0.132440	0.065747	0.083242	0.059638	0.060565	0.072663
	3 nodes, 8 branches	R2	0.93	0.90	0.90	0.90	0.90	0.62	0.91	0.85	0.92	0.92	0.88

Cl, chloride; K, potassium; K-IN, potassium infusion; Na, sodium; Na-IN, sodium infusion; R², coefficient of determination; RMSE, root mean square error; water-IN, water intake; water-OUT, urine volume.

Fig. 4

Ten-fold cross-validation of the training data.

(A) Linear regression model. (B) Support vector regression model. (C) Regression tree model.

The results of 10-fold cross-validation of the training data are shown. The linear regression model showed the highest accuracy, which corresponded to the root mean square deviation (RMSD) between the predicted and measured s-Na values of 1.11 mEq/L.

External validation

All training models were validated using the validation data from Nagoya University Hospital. Table 3 shows the external validation results with various combinations of 5–7 predictors. The linear regression model resulted in an RMSE of 0.0544170 and R² of 0.92 (Fig. 5A). The SVR model resulted in an RMSE of 0.0539600 and R² of 0.92 (Fig. 5B). The regression tree model resulted in an RMSE of 0.0593740 and R² of 0.91 (Fig. 5C). Among these three models, the SVR model showed the highest accuracy, which corresponded to the RMSD between the predicted and measured s-Na values was 1.62 mEq/L. The SVR model with all seven predictors had the best accuracy among all combinations of predictors.

Table 3

Differences in the prediction accuracy of external validation based on combinations of predictors

Predictors		Water-IN	●		●	●	●	●	●	●	●	●	●
		Na-IN	●	●		●	●	●	●	●	●		●
		K-IN	●	●	●		●	●	●	●			●
		Water-OUT	●	●	●	●		●	●	●	●	●	●
		s-Na	●	●	●	●	●		●	●	●	●	●
		s-K	●	●	●	●	●	●		●		●
		s-Cl	●	●	●	●	●	●	●		●	●
Modeling	Linear regression	RMSE	0.054417	0.055691	0.056090	0.054547	0.053714	0.111310	0.054492	0.054821	0.054608	0.055971	0.054813
		R2	0.92	0.92	0.92	0.92	0.93	0.68	0.92	0.92	0.92	0.92	0.92
	Support vector regression	RMSE	0.053960	0.055766	0.055585	0.053992	0.054335	0.128150	0.054364	0.054647	0.054215	0.055235	0.054584
	linear kernel c = 0.2337, γ = 0.0234	R2	0.92	0.92	0.92	0.92	0.92	0.57	0.92	0.92	0.92	0.92	0.92
	Regression tree model	RMSE	0.059374	0.068756	0.060310	0.060269	0.067869	0.095449	0.063258	0.080811	0.058820	0.057892	0.087652
	3 nodes, 8 branches	R2	0.91	0.88	0.91	0.91	0.88	0.76	0.90	0.83	0.91	0.91	0.80

Cl, chloride; K, potassium; K-IN, potassium infusion; Na, sodium; Na-IN, sodium infusion; R², coefficient of determination; RMSE, root mean square error; water-IN, water intake; water-OUT, urine volume.

Fig. 5

External validation of each model.

(A) Linear regression model. (B) Support vector regression (SVR) model. (C) Regression tree model.

The external validation results of the validation data are shown. The SVR model showed the highest accuracy, corresponding to the RMSD between the predicted and measured s-Na values of 1.62 mEq/L.

Discussion

In this study, we developed a machine learning model to predict s-Na at 6 h using clinically accessible parameters, such as s-Na, s-K, s-Cl, Na-IN, K-IN, water intake, and urine volume. According to the internal validation of accuracy with 10-fold cross-validation, the linear and SVR models showed high accuracy with RMSE of 0.037, and the RMSD between the predicted and measured s-Na values was 1.11 mEq/L. The external validation on these models further showed high accuracy with RMSE of 0.054 for the linear and SVR models, and the RMSD between the predicted and measured s-Na values was 1.62 mEq/L. Furthermore, the prediction accuracy did not significantly decrease even if one or two fewer predictors were missing, suggesting it would be applicable to cases in which some information are not available.

We selected several parameters as predictors in this study (Fig. 1A). Although other candidates as predictors, such as body weight and u-Na, may exist, these were not employed because they are not usually measured in the acute phases of treatment in clinical settings. Instead, we included not only s-Na, Na-IN, water intake, and urine volume but also s-K, s-Cl, and K-IN as predictors in this study, leading to the establishment of a predicting model for s-Na with high accuracy. K-IN is involved in the Adrogué–Madias formula [15], and a previous study showed that s-Cl is a good predictor for delayed hyponatremia after transsphenoidal surgery [16]. However, these results and ours do not necessarily imply that s-K, s-Cl, and K-IN have direct effects on s-Na, and explaining the results derived from artificial intelligence and machine learning models is sometimes difficult.

Hyponatremia is the most common electrolyte abnormality in general hospitalized patients. Even mild hyponatremia without apparent symptoms is associated with cognitive deficits [17], gait disturbance [18], and increased rates of falls and fractures [19-21]. Furthermore, hyponatremia is correlated with the increased mortality of patients with various diseases, such as lung cancer [22] and sepsis [23]; elderly patients [24]; and patients undergoing surgical intensive care [25]. Recent meta-analysis studies have confirmed that hyponatremia is a frequent presentation in up to 20% of hospitalized patients and is associated with longer hospital stays and readmissions [26]. However, selecting an appropriate treatment for hyponatremia and predicting s-Na response to that treatment is challenging, and patient-specific s-Na prediction algorithms are required. Our current s-Na prediction models could address this need and will help physicians, including those not specializing in endocrinology or nephrology, select an appropriate treatment.

Several studies have used machine learning methods to predict the onset of hyponatremia and for cluster classification of hyponatremia patients. Theerthagiri et al. used a multilayer perception and multivariate linear regression algorithm and have shown that the s-Na of patients with normonatremia could be predicted based on their history of health issues [13]. Voglis et al. have revealed that the machine learning model could predict the occurrence of hyponatremia after pituitary surgery, thus potentially reducing morbidity and improving patient safety [12]. Thongprayoon et al. have identified three clinically distinct phenotypes with differing mortality risks among a heterogeneous cohort of hospitalized patients with hyponatremia using an unsupervised machine learning approach [27]. However, no studies have provided predictions of the s-Na concentration of patients with hyponatremia based on the treatment choice. By using our s-Na predicting model developed in the current study, we can select an appropriate treatment for the safely correction of s-Na to prevent from a rapid increase in s-Na, which is the risk for ODS.

The Adrogué–Madias formula is a widely recognized clinical tool for calculating ΔNa after infusing 1 L based on three parameters: the amount of Na and K in a 1 L infusion and total body water [15]. However, its accuracy in predicting s-Na is limited to only when 1 L of fluid is administered, and the relationship between infusion volume and s-Na variability is not always linear [28]. In addition, measuring accurate total body water volume, which is included in input variables, is challenging in patients with severe hyponatremia in the acute phase when plasma AVP secretion and urine output are constantly changing, which could prevent physicians from using the Adrogué–Madias formula. In contrast, our current prediction model requires only clinically accessible and available parameters and thus could be more widely used to predict s-Na in clinical practice.

There are several limitations to this study. First, only severe cases were employed in this study so that accurate hourly data in intensive care electronic medical record systems were analyzed. Thus, the study data were obtained from a limited number of hospitals and did not include a large number of patients. Second, we excluded cases with u-Na <20 mEq/L and those with therapeutic interventions with DDAVP from the analysis because they have different pathology and may respond differently to the treatment. In cases of hyponatremia with dehydration, sufficient fluid supplementation with appropriate Na concentrations is needed. Strict s-Na monitoring is required for water intoxication since s-Na is likely to rise quite quickly. We could further address these cases by accumulating more data in future. Third, the predictors did not include information about the elapsed time since the onset of hyponatremia. While our model showed high accuracy in the prediction of s-Na, it is possible that responses of s-Na to treatment are different between the early and late stages of treatment. Fourth, this study focuses on the acute phase (6 h) of severe hyponatremia treatment and the prevention of rapid increase in s-Na. Therefore, predicting s-Na at 12 and 24 h from the initiation of the treatment would be possible if we use this model every 6 h with measured parameters. However, in a situation where measuring parameters every 6 h is difficult, as in this case, predictive models for s-Na at 12 or 24 h need to be established.

In conclusion, the predicted s-Na levels, using explainable machine learning algorithms and clinically accessible parameters, correlated well with the actual levels in the current study. Thus, our model could be applied to the treatment of hyponatremia in clinical practice.

Acknowledgments

This work was supported by the Hori Science and Art Foundation (date of approval: March 29, 2022).

Disclosure

None of the authors have any potential conflicts of interest associated with this research.

Some or all datasets generated during and/or analyzed during the current study are not publicly available but are available from the corresponding author on reasonable request.

This manuscript has been posted on a preprint server; Preprint with THE LANCET (https://ssrn.com/abstract=4422297, http://dx.doi.org/10.2139/ssrn.4422297).

Roles/Contributions

Conceptualization: TK, SO, DH, YA, HA. Data curation: TK, SO. Formal analysis: TK, SO. Funding acquisition: DH, HA. Investigation: TK, SO, DH, YA. Methodology: TK, SO. Project administration: SO, DH, HA. Resources: SO, DH, YA, HA. Software: TK, SO. Supervision: SO, DH, HA. Validation: TK. Visualization: TK. Writing: TK. Review and editing: TK, SO, DH, YA, HA.

References

• 1   Burst  V (2019) Etiology and Epidemiology of Hyponatremia. Front Horm Res 52: 24–35.
• 2   DeVita  MV,  Gardenswartz  MH,  Konecky  A,  Zabetakis  PM (1990) Incidence and etiology of hyponatremia in an intensive care unit. Clin Nephrol 34: 163–166.
• 3   Hawkins  RC (2003) Age and gender as risk factors for hyponatremia and hypernatremia. Clin Chim Acta 337: 169–172.
• 4   Upadhyay  A,  Jaber  BL,  Madias  NE (2006) Incidence and prevalence of hyponatremia. Am J Med 119: S30–S35.
• 5   Berghmans  T,  Paesmans  M,  Body  JJ (2000) A prospective study on hyponatraemia in medical cancer patients: epidemiology, aetiology and differential diagnosis. Support Care Cancer 8: 192–197.
• 6   Anderson  RJ,  Chung  HM,  Kluge  R,  Schirier  RW (1985) Hyponatremia: a prospective analysis of its epidemiology and the pathogenetic role of vasopressin. Ann Intern Med 102: 164–168.
• 7   Fenske  W,  Maier  SK,  Blechschmidt  A,  Allolio  B,  Störk  S (2010) Utility and limitations of the traditional diagnostic approach to hyponatremia: a diagnostic study. Am J Med 123: 652–657.
• 8   Mocan  M,  Terheș  LM,  Blaga  SN (2016) Difficulties in the diagnosis and management of hyponatremia. Clujul Med 89: 464–469.
• 9   Singh  TD,  Fugate  JE,  Rabinstein  AA (2014) Central pontine and extrapontine myelinolysis: a systematic review. Eur J Neurol 21: 1443–1450.
• 10   Verbalis  JG,  Goldsmith  SR,  Greenberg  A,  Korzelius  C,  Schrier  RW, et al. (2013) Diagnosis, evaluation, and treatment of hyponatremia: expert panel recommendations. Am J Med 126: S1–S42.
• 11   Lambeck  J,  Hieber  M,  Dreßing  A,  Niesen  WD (2019) Central pontine myelinosis and osmotic demyelination syndrome. Dtsch Ärztebl Int 116: 600–606.
• 12   Voglis  S,  van Niftrik  CHB,  Staartjes  VE,  Brandi  G,  Tschopp  O, et al. (2020) Feasibility of machine learning based predictive modelling of postoperative hyponatremia after pituitary surgery. Pituitary 23: 543–551.
• 13   Theerthagiri  P (2021) Forecasting hyponatremia in hospitalized patients using multilayer perceptron and multivariate linear regression techniques. Concurrency Computat Pract Exper 33: e6248.
• 14   Perianayagam  A,  Sterns  RH,  Silver  SM,  Grieff  M,  Mayo  R, et al. (2008) DDAVP is effective in preventing and reversing inadvertent overcorrection of hyponatremia. Clin J Am Soc Nephrol 3: 331–336.
• 15   Adrogué  HJ,  Madias  NE (2000) Hyponatremia. N Engl J Med 342: 1581–1589.
• 16   Fuse  Y,  Takeuchi  K,  Nishiwaki  H,  Imaizumi  T,  Nagata  Y, et al. (2023) Machine learning models predict delayed hyponatremia post-transsphenoidal surgery using clinically available features. Pituitary 26: 237–249.
• 17   Chung  MC,  Yu  TM,  Shu  KH,  Wu  MJ,  Chang  CH, et al. (2017) Hyponatremia and increased risk of dementia: a population-based retrospective cohort study. PLoS One 12: e0178977.
• 18   Fujisawa  H,  Sugimura  Y,  Takagi  H,  Mizoguchi  H,  Takeuchi  H, et al. (2016) Chronic hyponatremia causes neurologic and psychologic impairments. J Am Soc Nephrol 27: 766–780.
• 19   Renneboog  B,  Musch  W,  Vandemergel  X,  Manto  MU,  Decaux  G (2006) Mild chronic hyponatremia is associated with falls, unsteadiness, and attention deficits. Am J Med 119: 71.e1–71.e8.
• 20   Kuo  SCH,  Kuo  PJ,  Rau  CS,  Wu  SC,  Hsu  SY, et al. (2017) Hyponatremia is associated with worse outcomes from fall injuries in the elderly. Int J Environ Res Public Health 14: 460.
• 21   Barsony  J,  Kleess  L,  Verbalis  JG (2019) Hyponatremia is linked to bone loss, osteoporosis, fragility and bone fractures. Front Horm Res 52: 49–60.
• 22   Bartalis  E,  Gergics  M,  Tinusz  B,  Földi  M,  Kiss  S, et al. (2021) Prevalence and prognostic significance of hyponatremia in patients with lung cancer: systematic review and meta-analysis. Front Med (Lausanne) 8: 671951.
• 23   Castello  LM,  Gavelli  F,  Baldrighi  M,  Salmi  L,  Mearelli  F, et al. (2021) Hypernatremia and moderate-to-severe hyponatremia are independent predictors of mortality in septic patients at emergency department presentation: a sub-group analysis of the need-speed trial. Eur J Intern Med 83: 21–27.
• 24   Ioannou  P,  Panagiotakis  S,  Tsagkaraki  E,  Tsioutis  C,  Fragkiadakis  K, et al. (2021) Increased mortality in elderly patients admitted with hyponatremia: a prospective cohort study. J Clin Med 10: 3059.
• 25   Marshall  DC,  Salciccioli  JD,  Goodson  RJ,  Pimentel  MA,  Sun  KY, et al. (2017) The association between sodium fluctuations and mortality in surgical patients requiring intensive care. J Crit Care 40: 63–68.
• 26   Corona  G,  Giuliani  C,  Parenti  G,  Colombo  GL,  Sforza  A, et al. (2016) The economic burden of hyponatremia: systematic review and meta-analysis. Am J Med 129: 823–835.e4.
• 27   Thongprayoon  C,  Hansrivijit  P,  Mao  MA,  Vaitla  PK,  Kattah  AG, et al. (2021) Machine learning consensus clustering of hospitalized patients with admission hyponatremia. Diseases 9: 54.
• 28   Chen  S,  Shieh  M,  Chiaramonte  R,  Shey  J (2021) Improving on the Adrogue-Madias formula. Kidney360 2: 365–370.