Retrospective cohort study on the cross-sectional and longitudinal association between renal function and risk factors (Figure 2). Delphi technique (31–33) was used as the consensus methodology to construct a symptom-based classification framework for diabetes that comprised of symptom clusters. Explanatory power and complexity of statistical models with different combinations of biochemical, epidemiological, and symptom variables were compared by information criteria used in machine learning.

Three hundred two diabetes subjects aged 18 or above on antidiabetic medication were recruited consecutively from general outpatient clinics of public hospitals in Hong Kong through a 30-min health service programme from July 2014 to December 2019. Although there is no consensus on sample size calculation for comparing model quality, modeling bias is shown to be small and independent to sampling design empirically when sample size is over 100 (34).

Ten conventional medicine and CM dually trained experts in internal medicine and classical CM theory (Supplementary Material 6) were purposively sampled to join the Delphi consensus panel to diversify the schools of theory on symptom-based diabetes diagnosis.

Ninety-one symptoms, signs and behavioral factors (Supplementary Material 5) were extracted from existing literature. The presence of symptoms were prospectively collected face-to-face by physicians based on a standardized, structured, and field-tested questionnaire through the service program (35). Presence of symptoms was quantified by the days of presence within 2 weeks of consultation. Diagnosis of the 11 SSs was based on the presence of symptoms with reference to the symptom-based diagnosis framework derived from the expert consensus described below. To ensure data consistency, only 2 trained and synchronized physicians collected the data.

Biochemistry and demographics (listed in Data Analysis) of patients were retrieved from linked electronic medical record including a comprehensive and standardized annual multidisciplinary Risk Assessment and Management Programme (36). The Programme captures patients' demographics assessed by trained nurse, and the continuous change of key biochemical parameters. All outcome assessors were blinded from the analysis plan and biochemistry data to avoid bias. Longitudinal data was retrieved from the electronic medical system until 1 December 2020 by 2 researchers (K.W.C., K.Y.Y.) independently with a standardized form.

Consensus methods are commonly used in constructing agreeable frameworks of existing practice in medicine (31). Delphi technique, which involve iterative anonymous rating, has advantage in handling dominant opinion and geographical constraint (32). The consensus panel was formed based on their academic capacity, peer reputation, geographical coverage, and school of thoughts to maximize the coverage of opinion while being specific to diabetes. A panel of 10 was recruited to maximize participation and consensus efficiency (32, 37). The consensus result was used to reduce the dimension of symptom analysis from individual symptoms to symptom clusters.

The Delphi consensus was divided into 2 rounds. A questionnaire with Likert scale of 1–9 was designed as the medium for rating (Supplementary Material 7). Since CM has a rich history of using symptom-based diagnosis (23, 24), we sought to expand the scope of symptom analysis by extracting CM literature. Ninety-one baseline symptoms were extracted from all 11 documented SSs from key symptom-based diabetes guideline published by academy, academic society and government regulatory bodies (Supplementary Material 2) (38–41), and standardized by a national standard of medical terminology (42). The baseline questionnaire contained 300 symptoms and could be completed in 30 min in field tests.

In the first round, experts rated anonymously the importance of the symptoms in diagnosing each SS and were invited to supplement necessary symptoms. Consensus was pre-defined as two-third majority rating 1–3 (agreed as insignificant) or 7–9 (agreed as significant). Symptoms reached consensus were then removed from the questionnaire. The result in the first round was disseminated back to and confirmed by the consensus panel. Dissonant symptoms were circulated again in the second round along with the additional symptoms collected in the first round.

Symptoms reached consensus in either rounds were defined as core criteria for the diagnosis of the corresponding SS. Sensitivity analyses with different consensus definition (average score of 6.7) and handling of missing values (last-observation-carry-forward and extreme values) were conducted. Symptoms reached consensus in sensitivity analyses were defined as supplementary criteria. Results were confirmed by panel members.

Estimated GFR was calculated by the CKD-EPI equation (43). Education (no formal/primary/secondary/tertiary), type of diabetes (type 1/type 2/gestational), obesity (optimal weight/underweight/over-weight/moderately to severely obese), and smoking (non-/ex-/current smoker) and alcohol (non-/ex-/social/current drinker) history were further stratified. Univariable and backward stepwise multivariable regression models on the correlation between GFR and binary presence of individual symptoms were evaluated. Missing values were replaced by multiple imputation with 20 sets of data using age as auxiliary variable for the cross-sectional analysis.

Annual rate of GFR change over the follow-up period (in weeks) with repeated measures was analyzed by mixed models with unstructured covariance and adjusted for (1) baseline levels of hemoglobin A1c (HbA1c), systolic blood pressure, low-density lipoprotein, log-transformed urine albumin-to-creatinine ratio (UACR), and GFR, (2) baseline presence of symptoms or SSs, (3) gender, type of diabetes, smoking history, alcohol consumption, and obesity, (4) interaction between baseline GFR and visit, (5) and interaction between presence of symptoms or SSs and visit. The included biochemical parameters and epidemiological factors are known predictors of renal function decline. Baseline presence of symptoms was used as the explanatory variable over continuous change of symptoms as the analyses were focused on prediction instead of continuous association. The slope of change was analyzed with a linear growth model with sensitivity analysis of using repeated measures of biochemical parameters and blood pressures as covariates instead of baseline values, as previous studies showed that the majority of GFR decline is linear (44).

Five multivariable regression models were built using GFR as the dependent variable to assess the power of explaining the variance in GFR, namely (1) Individual Symptom-based Model, (2) Traditional CM Model (with SSs), (3) Traditional CM-Epidemiology Model (with SSs and epidemiological variables), (4) Conventional Medicine Model (with biochemical and epidemiological variables), and (5) Integrative Chinese and Conventional Medicine Model (with SSs, biochemical and epidemiological variables) (Table 1). Both cross-sectional regression models and longitudinal mixed models were evaluated.

Quality of the statistical models was assessed by Akaike information criterion (AIC) and Bayesian information criterion (BIC). AIC and BIC assessed the quality of statistical models in predicting GFR with different discount on model complexity, and are commonly used as model selection criteria in machine learning (45). Lower AIC and BIC refers to model that explain the variance better with less complexity, and is regarded as a better model. GFR was selected as the dependent variable for the models as diabetic kidney disease (DKD) is the most prevalent and costly complication of diabetes with limited known progression risk factors (44). Prevention of renal function decline has been suggested as the top priority among diabetic patients in our previous focus group interview series (25). The efficiency of diagnostic models in predicting rate of renal function has clinical significance (46). STATA 15.1 was used for analysis.

Majority of the 302 patients had type 2 diabetes, were non-smoker, centrally obese, non-insulin user and complicated with dyslipidemia, stage 2 CKD, and microalbuminuria (3.0 ± 1.1 mg/mmol) (Table 2). The mean age, GFR and HbA1c were 59.7 ± 0.7 years, 77.3 ± 1.5 ml/min/1.73m² and 7.7 ± 0.1%, respectively. Most presented symptoms and signs were susceptible to infections (80.8%), nocturnal polyuria (67.4%), forgetfulness (65.6%), past frequent intake of high fat diet (59.6%), malar flush (58.3%), malaise (56.3%) knee buckling (54.6%), lumbago (54.0%), dry skin (52.3%), and fatigue (49.7%) (Supplementary Material 6). Eight patients' record were irretrievable due to missing personal identifier. Median follow-up period was 220 weeks (4.2 years) [IQR: 212–234]. Median change of GFR was −2.3 ml/min/1.73m² [IQR: −8.7 to 1.4] over the period. 286 (94.7%) and 280 (92.7%) patients had annual GFR measurements until the 3rd and 4th year of follow-up.

The association between GFR and individual symptoms are summarized in Table 3. The frequency of nocturnal polyuria, defined as the reported average frequency of urination during sleep, demonstrated a strong inverse dose-dependent response with GFR (Supplementary Material 3) in the univariable analysis. In the multivariable symptom-based regression model, frequent urination (−9.1 ml/min/1.73m², 95%CI: −15.7 to −2.5, p = 0.007), nocturnal polyuria (−12.1 ml/min/1.73m², 95%CI: −17.7 to −6.5, p < 0.001), darkened complexion (−9.7 ml/min/1.73m², 95%CI: −17.6 to −1.8, p = 0.017), dark lips (−10.7 ml/min/1.73m², 95%CI: −16.5 to −4.8, p < 0.001) and dull purple color on lower limb (−13.5 ml/min/1.73m², 95%CI: −21.8 to −5.1, p = 0.002) are associated with reduced renal function after adjusting the presence of other symptoms. Further adjustment with biochemical parameters and demographics showed that edema (−7.1 ml/min/1.73m², 95%CI: −12.1 to −2.2, p = 0.005) and crenated tongue (−4.2 ml/min/1.73m², 95%CI: −8.2 to −0.3, p = 0.037) are independently associated with reduced GFR.

The 11 SSs extracted from literature described diabetes patients that co-present with: (1) SS1: lower gastrointestinal symptoms and obesity; (2) SS2: lower gastrointestinal symptoms and malaise; (3) SS3: upper gastrointestinal symptoms that related to emotion; (4) SS4: both upper and lower gastrointestinal symptoms; (5) SS5: upper gastrointestinal symptoms and fatigue/malaise; (6) SS6: upper gastrointestinal symptoms and cold extremities; (7) SS7: general dry and hot sensation; (8) SS8: dry sensation and fatigue/malaise; (9) SS9: waist and knee musculoskeletal and otorhinolaryngological symptoms; (10) SS10: general weakness, waist and knee musculoskeletal symptoms, nocturia/frothy urine, and hyposexuality; and (11) SS11: signs of poor circulation (Supplementary Material 1).

Invitation was sent to 11 experts. One refused due to work engagement and 10 experts joined the consensus panel (Supplementary Material 6). After 2 rounds of consensus, 90 symptoms reached consensus in total for the 11 SSs (Table 4). The inter-rater Cohen's Kappa between consensus definition of two-third majority and average score of 6.7 was 0.90 overall (0.89 and 0.90 for the first and second round of consensus, respectively). Sensitivity analyses showed that 39 symptoms reached consensus with consensus definition of 6.7 average score or after imputation of missing values. They were listed as supplementary criteria that add likelihood for the classification of the corresponding SS (Table 4).

The AIC and BIC of the 5 models are presented in Table 1. The Integrative Chinese and Conventional Medicine Model had the lowest AIC (4,478 vs. 5,824 vs. 4,966 vs. 7,926) and BIC (4,597 vs. 5,870 vs. 5,065 vs. 8,026) compared to the Traditional CM, Traditional CM-Epidemiology, and Conventional Medicine models, respectively. Using SS as the explanatory variable in statistical models resulted in better model quality with lower AIC and BIC when compared to the models using individual symptoms.

From the multivariable regression model of Integrative Chinese and Conventional Medicine, the presence of SS8 (fatigue, malaise, dry mouth, dry throat, weak pulse) is associated −1.1 ml/min/1.73m², 95%CI: −1.9 to −0.2, p = 0.011 more annual GFR drop after adjusting key biochemical parameters and demographics. In the symptom-based multivariable model using individual symptoms instead of SSs, edema (−1.8 ml/min/1.73m², 95%CI: −2.5 to −1.2, p < 0.001), epigastric bloating (−0.8 ml/min/1.73m², 95%CI: −1.4 to −0.2, p = 0.014), alternating dry and loose stool (−1.1 ml/min/1.73m², 95%CI: −1.9 to −0.4, p = 0.004), thin tongue (−3.0 ml/min/1.73m², 95%CI: −4.5 to −1.4, p < 0.001), deep pulse (−1.5 ml/min/1.73m², 95%CI: −2.2 to −0.9, p < 0.001), and unsmooth pulse (−1.2 ml/min/1.73m², 95%CI: −2.0 to −0.5, p = 0.002) were significantly and independently associated with faster GFR decline (Figure 3, Table 5). Result is robust in sensitivity analyses using repeated measures of biochemical parameters and blood pressures as covariates for adjustment.

Symptoms developed from clustering of pathophysiology (9, 13, 20) and have been used in disease monitoring, or as a treatment response predictor (7, 14). Distinctive clustering of frequent urination, excessive thirst, extreme hunger and unusual weight loss has characterized diabetes since 500 B.C.. Increased fatigue, irritability and blurry vision were supplemented since then. These symptoms have also been demonstrated to correlate with diabetes onset among non-diabetic population in the SHIELD study (21).

Independent associations have been established between symptoms and CKD. For instance, dermatological changes (e.g., pruritus, xerosis) were shown to correlate with uricemia (48). DKD was also defined as a composite of clinical presentations (edema, heart failure), epidemiology (long standing diabetes, hypertension), histopathology (nodular glomerulosclerosis, thickened basement membranes), and biochemistry (increased urine albumin) by Paul Kimmelstiel in 1936 (49). Symptoms that associate with the renal progression could further stratify patients for personalized management.

Currently, there is no standardized symptom-based stratification framework for diabetes (50). We identified discrepancies in the listing of commonly presented SSs and the related symptoms across different guidelines. Our newly synchronized framework is more concise. For instance, the defining symptoms of SS1 were simplified from 31 possible symptoms to 9 key symptoms (Supplementary Material 1). This facilitates the comparison of utility between individual symptoms and symptom clusters.

Infection susceptibility, forgetfulness, lumbago, knee buckling, and malar flush were commonly presented in our cohort. Classical presentation of frequent thirst and excessive hunger was less frequent, likely because the patients were already on hypoglycaemic agents that alleviated these symptoms.

From the regression analysis without adjustment of biochemical and epidemiological factors, a strong inverse dose-dependent relationship between GFR and nocturnal polyuria was observed, indicating a quantifiable assessment on nocturnal polyuria in clinical practice. Frequent urination, nocturnal polyuria, darkened complexion, dark lips, and dull purple color on lower limb are associated with lower GFR, which is commonly observed in practice.

After adjusting biochemical parameters and demographics including age, gender, HbA1c, and UACR, a set of different symptoms (edema, crenated tongue) were associated with poorer renal function independently. This is due to the collinearity between symptoms and biochemical parameters (e.g., frequent urination and UACR).

Using symptoms from the unadjusted symptom-based model can estimate renal function in resources limited setting or in mass population screening without requiring laboratory investigations. Symptoms from the adjusted model, meanwhile, can increase the explanatory power on top of known risk factors and indicated pathophysiology independent to the control of glucose, blood pressure, lipid, and albuminuria.

Model quality assessment showed that using symptom clusters provides better quality in modeling renal function when compared to using individual symptoms. This is due to the natural collinearity of symptoms in which the local dependence between symptoms can be better analyzed by grouping symptoms with shared pathophysiology and predictive power on renal function (51).

From the 4-year follow-up cohort, symptom clusters provide additional explanatory power on top of biochemical and epidemiological factors in stratifying the decline of renal function. Early diabetes presented with SS8, which included fatigue, malaise, dry mouth and dry throat, were associated with 1.1 ml/min/1.73m² faster renal function decline annually.

The presence of epigastric bloating, and alternating dry and loose stool are associated with faster GFR decline after adjusting blood glucose, blood pressure, lipid and urine albumin. From the 11 SSs that documented in CM literature, diabetes was stratified into three main groups (Figure 1): (1) SS1–SS6: with gastrointestinal presentation, (2) SS7, SS8: with classical diabetes presentations, and (3) SS9–SS11: with neurological and vascular presentations. The symptom-based stratification of diabetes depends heavily (6 of 11 SSs) on gastrointestinal presentations.

The symptom analysis and expert consensus converged to suggest an important role of gastrointestinal involvement in the renal progression of diabetes. Previous studies showed that gut microbiome is involved in the inflammatory and immune response in metabolic diseases, CKD and related phenotypical changes (52, 53). We previously demonstrated in vivo that toll-like receptor signaling pathway regulate DKD progression (54–56). Activation of the pathway via Paneth cells in small intestine due to microbiota dysbiosis can be a possible mechanism warrants further study (57). Endotoxin (e.g., lipopolysaccharide) translocation to systemic circulation due to disruption of gut barrier can also accelerate CKD progression (53).

Symptom-based research offers a new perspective in assessing correlations in pathogenesis among diseases, accelerating the discovery of secondary application of existing drugs (29). In this study, edema is a strongly independent cross-sectional and longitudinal predictor of reduced renal function. Further multi-level omics analysis stratifying the presence of edema with matching GFR and other confounders could reveal novel mechanisms in renal function decline among early diabetic patients.

Key challenge in symptom-based research is the standardization of subjective symptoms (9). Although the unharmonized definition of symptoms is likely to introduce random measurement error instead of systematic bias, and is unlikely to increase type I error in assessing model quality, a large sample size is needed to demonstrate statistical significance. In this study, thin tongue, deep pulse, and unsmooth pulse were strongly associated with faster renal function decline. Although these symptoms have long been suggested as signs of reduced kidney function and poor circulation in CM theory, standardized, and digitalized measurement is lacking until the past decade (58–60). Recent advancement in phenotype-oriented network in multi-omics data provided a platform to link up phenotype, genotype, omics, and drug targets for the integration of clinical presentation, genetics, and epigenetics (5, 9, 50, 61, 62). These advancements enhance the clinical translation and generalizability of symptom-based diagnosis research to the other parts of the world. Nevertheless, the description of pulse would require substantial work on standardization before being widely utilized.

Our cohort was conveniently sampled from public general outpatient clinics. Although the sampling is not random, selection bias was minimized by consecutive sampling. Majority of the diabetes patients in Hong Kong have follow-ups in public general outpatient clinics. The demographics of our cohort matched the territory-wide demographics (47). Besides, there is no standardized questionnaire for the assessment of symptoms internationally. We developed a data entry form for this purpose with field tests, and only recorded the presence of symptoms to minimize the error due to physicians' judgment. Error in symptom assessment is likely random error instead of systematic bias as the assessors were blinded from data analysis plan and biochemistry data. The measurement error is unlikely to increase type I error in assessing model quality. Furthermore, only a small group of experts with outstanding and diversified academic and clinical track record and was purposively sampled to form the consensus panel for better participation and coordination, which common for Delphi technique (32). Further large-scale untargeted mining by latent class analysis can identify the natural clustering of symptoms. In the analysis, the prescription was not analyzed as the control of blood glucose, pressure, lipid, and urine albumin were reflected by the change of HbA1c, systolic blood pressure, low-density lipoprotein, and UACR. Lastly, this study was designed to compare the quality of model performance and was not powered to evaluate the longitudinal predictiveness of each SS. The SSs that did not reach statistical significance are subject to type II error. Further large-scale correlation study is needed to precisely assess the clinical and statistical significance of each SS.