Development of an artificial intelligence-based early diagnostic system for light-chain amyloidosis

Abstract

Light-chain (AL) amyloidosis is a rare disease and its early diagnosis remains challenging. This study aimed to develop an artificial intelligence (AI)-based diagnostic assistance system to improve the early diagnosis of AL amyloidosis and facilitate earlier and more precise disease management. Through cooperation with 18 hospitals in the Chinese Registration Network for Light-chain Amyloidosis (CRNLA), 1,355 patients with AL amyloidosis were registered and followed up from January 2010 to January 2022. Ten variables that are easily monitored in clinics, including age, cardiac troponin I (cTnl), N-terminal prohormone B-type natriuretic peptide (NT-ProBNP), creatinine (Crea), albumin (ALB), total bilirubin (Tbil), alkaline phosphatase (ALP), interventricular septum (IVS), left ventricular posterior wall (LVPW), and ejection fraction (EF), were selected. An early assistant diagnostic model of AL amyloidosis was established using gradient boosting decision tree (GBDT), support vector machine (SVM), random forest (RF), and ensemble learning models. A validation subset of challenging cases from the external validation set was used to compare the performance of the AI-based early diagnostic assistance system to that of physicians. All of the typical machine learning algorithms, namely the GBDT, SVM, RF, and ensemble learning models, performed well. The ensemble learning model was considered to have the best performance in the external validation set based on its highest F1 value (0.94) and an area under the receiver operating characteristic (ROC) curve of 0.9784 (95% confidence interval [CI]:0.963–0.987). Performance comparisons between the AI-based early diagnostic assistance model and physicians suggested that the diagnostic accuracy of the AI model (0.95,95%CI:0.92–0.98;P < 0.05) substantially exceeded that of the hematologists (0.73,95%CI:0.7–0.82) and other physicians (0.65,95%CI: 0.61–0.71). The model established by AI, which uses routine laboratory and echocardiography results, can predict the possibility of AL amyloidosis in various scenarios, which may support the early diagnosis of this rare disease.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00277-025-06674-7.

Introduction

Light-chain (AL) amyloidosis is a rare disease that is associated with multiple organs and has a high mortality rate [1–3]. Globally, the early diagnosis of systemic AL amyloidosis remains challenging [4, 5]. Misdiagnoses or delays in diagnosis are even more common in highly populated and developing nations, such as China, Southeast Asia, South America, and Africa [6]. This is partially due to low awareness of this rare disease, and patients with AL amyloidosis are often found in many types of outpatient departments; however, large-scale standardized training for all physicians in clinics and hospitals with multiple specialties (cardiology, nephrology, neurology, etc.) is not feasible. Another reason for the difficulties in early AL amyloidosis diagnosis is the presence of nonspecific symptoms that often mimic other diseases, and indicators for an early diagnosis are currently lacking. A biopsy of the affected tissue, along with Congo red staining and electron microscopy, is the gold standard for diagnosing light-chain amyloidosis [7]; however, these tests are usually difficult to implement. Moreover, fat biopsy positivity is generally low [8], and renal and subendocardial biopsies are considered risky procedures. Therefore, physicians do not usually perform invasive examinations in patients exhibiting the early symptoms of AL amyloidosis. Therefore, it is necessary to establish a safe, economical, and feasible early diagnostic system for AL amyloidosis.

In recent years, there have been continuous landmark events in the various fields of artificial intelligence (AI). Due to the particularity of rare diseases, such as the small number of cases, the application of AI is not extensive. In this study, we aimed to develop an AI model to predict early AL amyloidosis based on laboratory and ultrasound results. With the help of AI technology, patients with AL amyloidosis can be diagnosed earlier and treated sooner, thereby prolonging their survival and improving their quality of life.

Methods

Ethics statements

The Medical Ethics Committee of Shengjing Hospital, China Medical University approved the present study (2020PS055J), which was performed in accordance with the principles of the Declaration of Helsinki. The requirement for informed consent was waived by the Medical Ethics Committee as electronic medical records were reviewed anonymously and retrospectively.

Study population

A sample set containing general information records of 1,355 AL amyloidosis patients, including complete blood count (CBC), basic metabolic profile (BMP), and echocardiography findings, was collected from 18 hospitals in the nationwide Chinese Registration Network for Light-chain Amyloidosis (CRNLA) from January 2010 to January 2022. AL amyloidosis was diagnosed based on tissue biopsies of amyloid deposits formed by immunoglobulin light chains [9]. Furthermore, 6,000 records of patients with non-AL amyloidosis were obtained from the Departments of Cardiology (n = 1000), Nephrology (n = 1000), Gastroenterology (n = 1000), Respiratory (n = 1000), Immunology, and Rheumatology (n = 1000), and records of healthy individuals at the Medical Examination Center (n = 1000) were randomly selected. The sources and characteristics of patients are shown in Table 1. A flowchart of this process is shown in Fig. 1.

Table 1.

Case characteristics of subsets

Subsets	Variables	AL amyloidosis Median (range)	Non-AL amyloidosis, median(range)
			Cardiology	Respiratory	Nephrology	Gastroenterology	Rheumatology	Healthy
Training & internal validation	Cases	1052	160	160	160	160	160	410
	Age	59(20–85)	56(18–89)	60(20–81)	52(26–78)	60(16–81)	51(14–78)	55(14–85)
	cTnI (µg/L)	5.16(0–597)	6.88(0.06–690)	2.01(0.01–82.01)	1.12(0.06–54.64)	0.86(0.03–9.02)	0.56(0–6.95.95)	0.17(0–36.76.76)
	NT-ProBNP (pg/mL)	424.8(7–103277)	852.4(16.2–35000.2)	126.5(7.28–2134)	336(8.61–20495)	125(10.4–3727)	102.6(16.67–6561)	118(13.93–1390)
	ALP (U/L)	124(34.9–637)	98.2(76–116.6.6)	72.5(51.1–177)	70.2(57.5–112)	89.3(61–355.1.1)	68.4(59–254.4.4)	83.3(56.4–115.8.4.8)
	Crea (µmol/L)	106(5.5–921.35.5.35)	60.3(36.6–987)	52.2(26.6–284)	302.6(50–9685.5.5)	52.7(29.6–483)	61(48.9–667)	67.7(52–394)
	ALB (g/L)	29(7.4–61)	37(28–51.3.3)	44(26.5–54.1)	28.5(27.8–55.7)	31(17.2–48)	38(26.4–58.6)	32(20.5–51.3)
	Tbil (µmol/L)	14.1(1–410.7.7)	10.2(3.5–31.9)	12.1(4.9–53.2)	11.5(3.2–36.6)	27.5(7.7–295.9.7.9)	11.9(3.2–35.5)	12.8(3.4–122.5.4.5)
	IVS (mm)	12.6(1–21)	12.8(2.1–20)	8.5(2.2–14)	6.6(3.8–14)	5.7(2.4–13)	8.2(2.5–14)	8(2.1–14)
	LVPW (mm)	11.8(0.9–23)	12.3(1.9–22)	8.5(2.1–13)	6.4(2.1–11.8)	7.7(2.4–12)	8.9(1.0–13)	7.3(2.2–11)
	EF (%)	60.7(21–88)	55(22–65)	65(39–67)	69(33–75)	66(31–72)	65(54–71)	62(42–69)
External validation	Cases No.	469	60	60	60	60	60	103
	Age	61(31–84)	54(16–84)	58(25–80)	53(22–81)	58(15–75)	51(16–72)	54(14–81)
	cTnI (µg/L)	5.3(0–961.8.8)	6.65(0.02–780)	0.26(0.01–31.01)	0.33(0.01–9.5)	0.115(0.01–17.7)	0.089(0–5.35.35)	0.112(0.01–4.16)
	NT-ProBNP (pg/mL)	590.8(6–35000)	786(24.19–35000)	254(43.4–6556)	328(23.66–2809)	163(47.3–3188)	112.3(13.47–9685)	101.5(27.75–952)
	ALP (U/L)	132(26–554)	95.4(56–116.6.6)	75.4(55.2–252)	71(44.3–140.5.3.5)	90.6(46.2–428.2.2.2)	68(34–150.4.4)	80(36.6–160.21.6.21)
	Crea (µmol/L)	101.3(44–874)	73.5(39.9–344)	80.4(42.5–627)	335.6(31.1–8183)	86(53.4–289.2.4.2)	96.2(56.4–188)	65.5(52.5–254.1.5.1)
	ALB (g/L)	31(12.58–68.58)	41.5(30–55.7.7)	39.7(29.8–48.8)	31.8(22–45.7.7)	41.1(18.5–46.7)	46.8(29–52.4.4)	44.8(29–50.6.6)
	Tbil (µmol/L)	13.5(2.5–78.8)	17.4(5.8–91.4)	17(4.1–49.4)	26.8(6.9–40.2)	26.4(5.6–112.3.6.3)	14.6(5.2–48.8)	12.9(4.2–36.9)
	IVS (mm)	12.8(1.8–60)	11.6(2.7–18)	9.2(3.1–15)	7.7(2.4–14)	6.5(3.3–12)	7.2(4.5–16)	7.1(2.8–13)
	LVPW (mm)	12(1.7–40)	9.8(2.1–12)	6.7(3.5–11)	8.1(2.7–11)	8(3.5–11.5)	9.4(2.8–12)	8.6(1.9–13)
	EF (%)	60(25–80)	56(26–68)	59(31–74)	65(38–70)	62(45–71)	65(41–69)	66(40–68)

AL amyloidosis Light-chain amyloidosis; cTnI Cardiac troponin I; NT-ProBNP N-terminal-pro-hormone B-type natriuretic peptide; ALP Alkaline phosphatase; Crea Creatinine; ALB Albumin; Tbil Total bilirubin; IVS Interventricular septum; LVPW Left ventricular posterior wall; EF Ejection fraction

Fig. 1.

Study profile

Variable determination

Diagnostically significant variables were selected based on the clinical manifestations and characteristics of AL amyloidosis. To be selected, these variables had to be routine and easily obtainable, thus allowing the AI early assistance diagnostic model to be widely used in the future.

Data preprocessing

AL amyloidosis is a rare plasma cell disease with very few reported cases. Data collection involved multiple centers across the country and a total of 1,355 cases were identified. Managing missing values and preserving cases was critical during data processing. Therefore, the impact of different missing values on the accuracy of the machine models was compared and data with excessive missing values were removed. In addition, disturbance methods were used to fill in the missing values.

AI early assistance diagnostic model construction

The sample set was randomly split into training and internal validation sets at a ratio of 2:1. Three different AI algorithms were used: random forest (RF) [10], support vector machine (SVM) [11], and gradient boosting decision tree (GBDT) [12, 13]. RF is a combination of tree predictors that are less prone to overfitting than a single decision tree [12]. An SVM can project data into a high-dimensional feature space, in which a hyperplane can be identified and used for classification. The GBDT is an iterative decision tree algorithm composed of multiple decision trees, and the conclusions are based on all decision trees. An ensemble learning machine that integrates the above three classifiers was selected as the classifier. To compare the differences between traditional methods and AI, linear regression (LR) was used to analyze the data [14]. Precision refers to the possibility of being positive when predicted as positive (AL amyloidosis), and recall represents the true-positive rate, which is the proportion of positive cases that are correctly classified. The F1 score is the harmonic mean of the precision and recall. The area under the receiver operating characteristic curve (AUC) values of the receiver operating characteristic (ROC) curves [15], precision, recall, and F1 score were used to compare the accuracies of the different algorithms.

AI early assistance diagnostic model validation

To validate the AI model, 469 new AL amyloidosis patients from the CRNLA were assigned to the external validation set after the cutoff date of January 31, 2021. Patient characteristics are presented in Table 1. In addition, to compare human diagnostic accuracy with AI models, 100 physicians from 10 hospitals were invited to examine the 872 records of patients with AL amyloidosis and non-AL amyloidosis. These physicians were from the Department of Hematology (senior, intern, and junior) and from non-hematology departments (Cardiologists, Nephrologists, Gastroenterologists, and Other physicians). Furthermore, physicians from general hospitals and primary care centers were also present.

Statistical analyses

Breakdowns of the predictive labels with reference to the ground-truth labels were depicted as confusion matrices, which were used to calculate the accuracy, sensitivity, and specificity (Python version 3.7.0 [Python Software Foundation]). All of the experimental programs in this study were written in Python (Wilmington, DE, USA), and the interpreter used Python version 3.6. The AI development tool used in this study was Sciket-Learn version 0.20. The RF used in this study was based on Sklearn, the ensemble class was based on sciket-learn, the SVM was based on the SVM class of Sciket-Learn, and the GBDT was based on the gradient boosting classifier class of Sciket-Learn. Deep learning was based on TensorFlow 1.12 (San Francisco, CA). NumPy 1.15.4 was used to process and transform the arrays, and Matplotlib 3.0.2 was used to draw the ROC curves.

Results

Ten variables based On CBC, BMP, and cardiac ultrasound findings were selected

The intended use of our AI-based diagnostic assistance model is for early screening, and the selected variables should be rapidly obtained in hospitals and clinics. Based on this consideration, the CBC, BMP, and cardiac ultrasound results were consistent. Increased age is a major risk factor for AL amyloidosis [4]. Total bilirubin (Tbil) and alkaline phosphatase (ALP) are indicators of liver involvement in AL amyloidosis [16, 17]. Creatinine (Crea) and albumin (ALB) levels are also important indicators of renal involvement. Cardiac Troponin I (cTnl) and N-terminal prohormone B-type natriuretic peptide (NT-ProBNP) are associated with the degree of cardiac involvement [18]. Typical echocardiographic findings can be used as diagnostic indicators [19]. To alleviate interference from human identification factors, we selected three quantitative variables: interventricular septum (IVS), ejection fraction (EF), and left ventricular posterior wall (LVPW). Simultaneously, the Pearson index was applied to calculate the relevance. The correlation coefficients are as follows: 0.61 (LVPW) >0.51 (ALB) >0.37 (IVS) >0.22 (Crea) >0.13 (NT-ProBNP) >0.10 (EF) >0.04 (ALP) >0.03 (Age) >0.01 (cTnl) >0.009 (Tbil). Therefore, these ten variables were selected for the AI-based assistance diagnostic model.

Filling of missing values with random perturbation

In the original database, some values were missing, with the exception of age. First, we filled in the missing values with 0 and classified them using the GBDT, RF, SVM, and LR models. The performance of each model is presented in S1 Fig in the Supplement. We found that replacing missing values with 0 directly resulted in poor classification performance, as shown in S1 Fig A. Although some classifiers can achieve a favorable classification performance, they perform poorly on other classification algorithms, as shown in S1 Fig B and C. Although existing algorithms can perform well on the training set, when the sample set is small, there is a possibility of overfitting, as shown in S1 Fig E and F. Therefore, we selected data with up to three missing variables as the sampling data for model construction. Finally, we used 1,052 AL amyloidosis patients and 1,210 non-AL amyloidosis patients to develop the AI-based early assistance diagnostic model and divided the final data into training and internal validation sets in a 2:1 ratio.

Using 0 for all data may lead to a decrease in model resolution, as the contribution of each variable is ignored. To further increase the diversity of the data, we added random values to each missing value. This added value was determined using different probability perturbations based on the average value of each variable. We assumed that the normal value range of a missing variable was [a, b], where a is the lower limit and b is the upper limit. This value is a random number in the range , where α (0 < α < 1) is the disturbance factor [20]. Considering cTnl as an example, its value range is 0–0.2.2. When α is 0.5, the value replacing the missing data has a value range of 0.1 ± rand (0,0.05). Based on the above strategies, α was set as 1, 0.5, and 0.25. The classification performance of the model is illustrated in S2 Fig. From the graph, it can be observed that adding random disturbances can effectively improve the model’s performance under various classifiers, and selecting disturbance factors α to 0.25 achieves the best overall performance. Therefore, the missing values of the original data were replaced with random values, where the disturbance factor α was set as 0.25. A comparison of the standardized boxplots for each feature is displayed in S3 Fig. We found that each variable had a similar distribution, thus ensuring the generalization ability of the model.

Performance of the AI-based early assistance diagnostic model

After data preprocessing, 2,262 cases were divided into AL amyloidosis (n = 1,052) and non-AL amyloidosis (n = 1,210) cases, and 1,542 cases in the training set included 732 AL amyloidosis and 810 non-AL amyloidosis cases. The 720 cases in the internal validation set included 320 AL amyloidosis and 400 non-AL amyloidosis cases. In addition, 872 new cases were collected as the external validation set, including 469 AL amyloidosis and 400 non-AL amyloidosis cases. Three different AI algorithms with 10 variables were used: RF, SVM, and GBDT. Considering that the above sample set performed similarly in the RF, SVM, GBDT, and other classifiers, an ensemble learning machine integrating the above three classifiers was selected as the classifier, and its internal and external validation results are shown in Fig. 2; Table 2. The model performed well on both the internal and external validation sets.

Fig. 2.

Performance comparison of internal and external validation sets of artificial intelligence (AI)-based models for early assistant diagnosis of light-chain (AL) amyloidosis. A internal validation results, and (B) external validation results

Table 2.

Precision, recall, and the F1 score of the AI early assistant diagnosis models for AL amyloidosis with 10 variables

Subset	Disease	AI models	Precision	Recall	F1 score
Internal validation	AL amyloidosis	RF	0.92	0.87	0.90
		SVM	0.70	0.89	0.78
		GBDT	0.92	0.89	0.90
		Ensemble learning model	0.96	0.96	0.93
	Non-AL amyloidosis	RF	0.87	0.92	0.89
		SVM	0.87	0.65	0.74
		GBDT	0.89	0.92	0.91
		Ensemble learning model	0.89	0.89	0.93
External validation	AL amyloidosis	RF	0.94	0.90	0.92
		SVM	0.64	0.94	0.76
		GBDT	0.94	0.92	0.93
		Ensemble learning model	0.94	0.93	0.94
	Non-AL amyloidosis	RF	0.92	0.95	0.94
		SVM	0.92	0.58	0.71
		GBDT	0.93	0.95	0.94
		Ensemble learning model	0.93	0.94	0.92

RF Random forest; GBDT Gradient boosting decision tree; SVM Support vector machine

Performance comparison with AI-based early assistance diagnostic model and physicians

A total of 100 physicians from 10 hospitals in various regions of China were invited to examine the cases based on 10 variables. The hospitals from which these physicians came did not necessarily coincide with the hospitals from which the cases were collected. The constructed AI-based model and the LR method of traditional statistics were simultaneously used to interpret the same cases. Table 3 presents the results of the AI and physicians’ diagnostic identification. Figure 3 displays a performance comparison of the proposed ensemble learning model with physician identification and the LR model. We found that all of the physician’s identification accuracy rates were 68.5%. For the traditional LR method, the cumulative accuracy rate was 85.7%. In the proposed ensemble learning model, the cumulative accuracy rate was 92.7%, which exhibited the most accurate identification in comparison to the physicians and LR. Stratified analysis by different physicians indicated that hematologists made more accurate diagnostic predictions (73.2%) than non-hematologists (65.0%). Internal hematologists appeared to have higher accuracy (75.5%) than junior (69.3%) and senior (73.6%) hematologists. Physicians from general hospitals were more accurate (69.0%) than those from primary care hospitals (64.3%). Regardless of the group, the AI-based model exhibited good stability and high accuracy.

Table 3.

Performance of AI and physicians in diagnosing AL amyloidosis

	Accuracy (95%CI)	Sensitivity (95%CI)	Specificity (95%CI)
Ensemble learning model	0.95 (0.92 ~ 0.97)	0.93 (0.85 ~ 0.95)	0.94 (0.91 ~ 0.98)
Physicians
General hospitals
Hematology
Senior hematologists (n = 14)	73.6 (58.9–88.2)	68.4 (50.0–76.4.0.4)	84.1 (79.9–92.1)
Intern hematologists (n = 13)	75.5 (70.7–82.1)	75.4 (72.3–80.6)	75.7 (70.9–86.3)
Junior hematologists (n = 13)	69.3 (61.1–78.3)	77.1 (69.8–88.0)	57.9 (52.3–69.0)
Non-Hematology
Cardiologists (n = 10)	66.9 (62.9–79.6)	65.4 (61.9–76.3)	69.4 (62.7–78.2)
Nephrologists (n = 10)	65.4 (59.9–70.1)	60.5 (49.6–69.6)	73.5 (66.3–79.2)
Gastroenterologists (n = 10)	68.9 (62.2–78.3)	66.7 (60.5–72.4)	73.3 (69.9–79.4)
Other physicians (n = 10)	55.0 (48.7–66.1)	51.2 (42.7–58.6)	64.7 (55.3–72.9)
Primary hospitals
General practitioners (n = 20)	64.3 (58.8–72.4)	72.1 (65,7–79.0)	51.9 (45.5–59.8)

Fig. 3.

Performance comparison of the proposed ensemble learning model with physician identification and linear regression (LR) model. A diagnosis accuracy of AI-based model, the physicians and LR model. B diagnosis accuracy of AI-based model, hematologists, non-hematologists and LR model. C diagnosis accuracy of AI-based model, hematologists of different levels (junior, intern, senior) and LR model. D diagnosis accuracy of AI-based model, non-hematologists (cardiologists, nephrologists, gastroenterologists, and other physicians) and LR model. E diagnosis accuracy of AI-based model, physicians from general hospitals, primary hospitals, clinics and LR model

The variable reduction to fit different actual scenarios

Considering the issue of missing variables in actual scenarios, further variable compression was required to better satisfy the diagnostic needs of different scenarios. Based on the correlation coefficients between the 10 selected variables and the final diagnostic results, we arranged them in descending order as follows: LVPW (0.61) > ALB (0.51) > IVS (0.37) > Cr (0.22) > NTpro-BNP (0.13) > EF (0.10) > ALP (0.04) > Age (0.03) > cTnl (0.01) > Tbil (0.009). Variables were removed in descending order of relevance, including Tbil, cTnl, Age, ALP, EF, NTPro, Cr, IVS, ALB, and LVPW, and the AI-based model performance was evaluated separately. Figure 4 displays the AUC range analysis for different variable combinations. When certain variables are missing, this model can also achieve good classification performance, fulfilling the need for accurate diagnostic prediction in primary care centers with inadequate laboratory results, with or without echocardiography.

Fig. 4.

The AUC range analysis with different variables combinations. Considering the issue of missing variables in actual scenarios, further variable compression was required to better satisfy the diagnostic needs of different scenarios. Variables were removed in descending order of relevance, including Tbil, cTnl, Age, ALP, EF, NTPro, Cr, IVS, ALB, and LVPW, and the AI-based model performance was evaluated separately. In cases where certain variables are missing, this model also achieves satisfactory classification results

Discussion

Light-chain amyloidosis is a rare disorder of the plasma cells. In China, family doctors and community hospitals have virtually no experience in diagnosing AL amyloidosis. The low incidence rate and the lack of experience in grassroots hospitals can lead to a high rate of missed diagnoses and misdiagnoses in patients with AL amyloidosis. Only 28.2% of patients are diagnosed within 6 months of symptom onset, and more than 70% of patients experience delayed diagnosis [4, 21]. In addition, most patients require multiple visits to different medical specialists. A substantial proportion of patients (31.8%) have been found to visit five different physicians before being diagnosed with amyloidosis, and only 7.6% received a diagnosis after visiting one physician. The most common physician for the first visit was a primary care physician (64.9%). The second visit most commonly involved cardiologists (24.5%), hematologists/oncologists (16.5%), nephrologists (18.1%), and gastroenterologists (14.6%) [22, 23]. Patients usually receive a correct diagnosis from a hematologist/oncologist (34.1%), as these doctors have more experience in diagnosing AL amyloidosis in the past. Nephrologists and cardiologists were found to provide correct diagnoses for 22.6% and 18.7% of the patients, respectively [24]. The external validation results also indicated that hematologists had a higher diagnostic accuracy rate than other specialized internal medicine physicians. Nevertheless, difficulties in early diagnosis remain a significant obstacle with regards to a positive prognosis for patients with AL amyloidosis, and establishing an early and accurate diagnostic assistant system for AL amyloidosis remains a challenge worldwide.

AI is a rapidly evolving technology in many industries, and is mainly used to improve performance, precision, and time efficiency, and to reduce costs [25]. However, it is difficult to develop an AI-based model for rare diseases, such as AL amyloidosis, because of the lack of data. To solve this problem, we used the CRNLA, which was established in 2019 and is currently the largest AL amyloidosis database in China, to obtain a relatively large AL amyloidosis dataset. Second, we replaced missing values with a disturbance value to reduce data loss. In addition, ensemble learning was employed by combining a variety of learning models to improve the stability. Our study explored the application of AI in the field of rare diseases, which has important clinical significance.

By comparing the performances of the AI-based model and physicians, we discovered that the diagnostic sensitivity of the AI-based model substantially exceeded that of physicians. We believe that the low diagnostic accuracy of general practitioners in primary care centers is mainly related to the lack of awareness of this rare disease, which further reflects the necessity of the AI-based early assistance diagnostic model for early screening during the first consultation. Moreover, although the overall diagnostic prediction accuracy rate exceeded 50%, the survey was mainly based on a yes or no answer. In a complex real-world scenario, the awareness of this rare disease is more challenging, and the true diagnostic prediction accuracy rate would be much lower than that observed in this study. Overall, the stability of the AI-based model is excellent and can resolve the differences between different levels and specialties.

In this study, the variables used were CBC, BMP, and color Doppler ultrasonography, which are inexpensive and easy to perform. Two application scenarios were designed (Fig. 5). The first is passive screening, and the AI-based model can be implanted into the hospital information management system (HIS) as long as the corresponding variables are input automatically, suggesting that the probability of AL amyloidosis is greater than 90%; the system can warn physicians and prompt for further diagnosis. This tool can also generate a real-time early warning regardless of the department the patient visits. In addition, for active screening, when a diagnosis of AL amyloidosis is considered, physicians can apply the AI-based model for query, and early identification can be suggested for high-risk patients. Through this AI-based early diagnosis assistant model, we provide a new approach to solve the issues associated with early diagnosis of AL amyloidosis, allowing patients to receive early treatment, which is more conducive to improving organ function and improving their prognosis. Although AI has a high predictive accuracy, until now, further application of tissue biopsy to make a definitive diagnosis is still needed.

Fig. 5.

Clinical application scenarios of AI-based AL amyloidosis early assistant diagnostic system. Two application scenarios were designed. For passive screening, the AI-based model can be implanted into the hospital information management system (HIS) as long as the corresponding variables are input automatically, suggesting that the probability of AL amyloidosis is greater than 90%; the system can warn physicians and prompt for further diagnosis. For active screening, when a diagnosis of AL amyloidosis is considered, physicians can apply the AI-based model for query, and early identification can be suggested for high-risk patients

There were some limitations in the present study. First, this was a retrospective study; the selection bias was inevitable. Second, the indicators such as right-ventricular cardiac parameters, creatinine clearance rate and 24-hour urine protein are closely related to the diagnosis and prognosis of AL-type amyloidosis. They are also the common markers for both cardiac amyloidosis and renal amyloidosis. Due to the accessibility and early specificity of these indicators, they were not included in the design of this AI model. Although the AL model demonstrated high diagnostic efficacy in the current validation results, further verification is still needed in more and more complex cases. Third, AL amyloidosis is a rare disease. The database used in this study is the largest one for AL amyloidosis in China. The patient data were sourced from 18 hospitals across the country, and inevitably, data missing was encountered. In order to ensure the quality and quantity of the data, we conducted multiple calculations and adopted the method of random perturbation to fill in the missing values, aiming to minimize the impact on the AI model. Fourth, the current early diagnosis model for AL is still in the initial research stage and has a long way to go before it can be used in clinical settings. It involves numerous issues such as patent conversion, software development, system compatibility, and data privacy protection, and the road ahead is fraught with difficulties. Future larger and higher-quality studies are required to validate our results and to rectify the defects.

Conclusion

Using routine data as the breakthrough point, we adopted an AI-based method to propose the automatic integration of routine test results and cardiac echocardiography findings to establish an early diagnosis assistant model for AL amyloidosis, all in an effort to improve the rate of early diagnosis. We believe that our research may provide a new and effective way to study rare diseases and diagnose them early.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AL

Light-chain

AI

Artificial intelligence

CRNLA

Chinese Registration Network for Light-chain Amyloidosis

cTnl

Cardiac troponin I

NT-ProBNP

N-terminal prohormone B-type natriuretic peptide

Crea

Creatinine

ALB

Albumin

Tbil

Total bilirubin

ALP

Alkaline phosphatase

IVS

Interventricular septum

LVPW

Left ventricular posterior wall

EF

Ejection fraction

GBDT

Gradient boosting decision tree

SVM

Support vector machine

RF

Random forest

ROC

Receiver operating characteristic

CI

Confidence interval

CBC

Complete blood count

BMP

Basic metabolic profile

LR

Linear regression

HIS

Hospital information management system

Author contributions

Conceptualization: W. Y., H.W.Data curation: W.Y., X.L.Formal analysis: T.H.Funding acquisition: W.Y., J.C., H.W.Methodology: T.H., H.X., J.C., B.L.Software: T.H., J.C.Validation: L.Z., Y.L., C.S., X.W., L.G., F.J., B.L., C.L., Z. C.Writing-original draft: W.Y., H.W.Writing-review&editing: Y.W., J.L., J.C., H.W.All authors have read and agreed to the published version of the manuscript.

Funding

JC has received a scholarship from the National Natural Science Foundation of China (Grant Nos. 61972079 and 61772126), Shingling Liao Ning Revitalization Talents Program (grant no.: XLYC1802100), Key R&D Program of Liaoning Province (grant no.2019JH2/10100027). HW is supported by Natural Science Foundation of Liaoning Province (grant no.2022-YGJC-61;2022-MS-219). WY is supported by the 345 Talent Project in Shengjing Hospital of China Medical University.

Data availability

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

Declarations

Ethics statements

The studies involving human participants were reviewed and approved by the Medical Ethics Committee of the Shengjing Hospital, China Medical University. The requirement for informed consent was waived by the Medical Ethics Committee as electronic medical records were reviewed anonymously and retrospectively.

Competing interests

The authors declare no competing interests.