The effect of metabolic risk factors on the natural course of gastro-oesophageal reflux disease (GORD), which remains elusive, was quantified.

The population included 3669 subjects undergoing repeated upper endoscopy. Data were analysed using a three-state Markov model to estimate transition rates (according to the Los Angeles classification) regarding the natural course of the disease. Individual risk score together with the kinetic curve was derived by identifying significant factors responsible for the net force between progression and regression.

During three consecutive study periods, 12.2, 14.9 and 17.9% of subjects, respectively, progressed from non-erosive to erosive disease, whereas 42.5, 37.3 and 34.6%, respectively, regressed to the non-erosive stage. The annual transition rate from non-erosive to class A–B disease was 0.151 per person year (95% CI 0.136 to 0.165) and from class A–B to C–D was 0.079 per person year (95% CI 0.063 to 0.094). The regression rate from class A–B to non-erosive disease was 0.481 per person year (95% CI 0.425 to 0.536). Class C–D, however, appeared to be an absorbing state when not properly treated. Being male (relative risk (RR) 4.31; 95% CI 3.22 to 5.75), smoking (RR 1.20; 95% CI 1.03 to 1.39) or having metabolic syndrome (RR 1.75; 95% CI 1.29 to 2.38) independently increased the likelihood of progressing from a non-erosive to an erosive stage of disease and/or lowered the likelihood of disease regression. The short-term use of acid suppressants (RR 0.54; 95% CI 0.39 to 0.75) raised the likelihood of regression from erosive to non-erosive disease.

Intraoesophageal damage is a dynamic and migratory process in which the metabolic syndrome is associated with accelerated progression to or attenuated regression from erosive states. These findings have important implications for the design of effective prevention and screening strategies.

Our study was based on a voluntary health promotion programme at National Taiwan University Hospital (NTUH) that used a standard protocol including a physical examination, blood chemistries, plain radiography, abdominal ultrasonography and endoscopy. Most subjects were invited to undergo an upper gastrointestinal endoscopy annually. Such a scheme is confirmed to be effective for cancer prevention in areas where the upper gastrointestinal cancers are prevalent.^{–}

We enrolled patients who underwent at least two endoscopic examinations. Excluded were those who received proton pump inhibitors (PPIs) or histamine-2 receptor antagonists (H2RAs) in the 4 months preceding the first endoscopy, those who underwent gastrectomy and those with malignancy. National Health Insurance in Taiwan covers a 4-month course of treatment with a PPI or an H2RA for those who show signs of erosive oesophagitis or peptic ulcer disease after endoscopy.

Prior to the examination, a self-administered questionnaire was used to collect information on demographics, social habits and medical/medication histories. We defined symptoms of GORD as the presence of troublesome heartburn, acid regurgitation or both. Heartburn was defined as a burning sensation in the retrosternal area and acid regurgitation as the perception of flow of refluxed gastric contents into the mouth or hypopharynx. The frequency was once a week or more over the past 3 months. Self-reported data were confirmed in a face-to-face interview with an internist.

Participants were evaluated for metabolic risk factors, including measurements of body mass index (BMI), waist circumference, blood pressure, plasma glucose, total cholesterol, high-density lipoprotein (HDL) cholesterol, triglycerides and uric acid levels. According to the modified criteria for Asians,

After the evaluation of metabolic risk factors, subjects underwent endoscopy. Erosive oesophagitis was scored using the Los Angeles (LA) classification system with standard comparator photos.

We modelled GORD’s natural history as a three-state Markov process by defining state 1 as NE disease, state 2 as LA class A–B oesophagitis and state 3 as class C–D oesophagitis (_{1}) and from state 2 to 3 (λ_{3}), and two instantaneous regression rates from state 2 to 1 (λ_{2}) and state 3 to 2 (λ_{4}).

Cumulative risk for each transition was computed by transition probabilities that were a function of transition rates λ_{1}–λ_{4} and follow-up time by using the method of Chen

Time intervals between endoscopic examinations were recorded to build up a continuous-time Markov process for the three-state model. We estimated the transition rates labelled in ^{–}

We presented the model parameters derived from the complete data set of subjects. To test the predictive validity of the current model, we also performed cross-validation by splitting data into 2/3 for deriving the model and 1/3 for validation of the model. The observed transition histories in the validated data set were compared with the predicted ones that were computed by the application of parameters trained from the derived data set.

As we are interested in the effect of metabolic risk factors on transition rates, a univariate regression analysis was therefore done to assess the effect of each component on transition rates. The exponential Markov regression form was adapted to estimate relative risk (RR), which is done by taking exponentials of the regression coefficients of the Markov regression. Besides the metabolic profile, factors considered in the regression included basic demographic information, lifestyle factors, symptoms of reflux and the use of acid suppressants following screening. It should be noted that each predictor in the same individual may vary from time to time and was repeatedly evaluated along with each endoscopic screening. They are treated as time-varying covariates, which means that their contributions to progression and regression of GORD may depend on the status they had at the time preceding the next transition during a given epoch. Thus, the net force of dynamic change of each covariate contributing to progression and regression is worthy of investigation.

Take smoking status (smoking, X = 1; no smoking, X = 0), for example. An individual at time t_{0} was a current smoker (X = 1), he developed A–B during time interval (t_{1}–t_{0}), quit smoking, was treated as non-smoking (X = 0), at time t_{1}, regressed to NE during the time interval (t_{2}–t_{1}), and stayed as NE without smoking until time t_{3}. Thus, as smoking status changed with time, its effect in each epoch makes different contributions to disease progression and regression in the same individual. The net force of smoking on the state-to-state transitions can be considered in this manner to aggregate each individual change into a population-average net effect expressed by the difference of regression coefficients between progression and regression. The same phenomenon may be applied to PPI/H2RA use that is associated with progression on the grounds of indication and regression probably causally related through treatment.

Hypothesis testing for such net force for each risk factor mentioned above is performed as follows. The transition rate function is first developed:

λ_{j}_{0j}(_{j}X

where λ_{01} and λ_{02} are baseline progression and regression rates, β_{1} and β_{2} are the corresponding regression coefficients, and _{1}–β_{2} (net effect) = 0. The alternative hypothesis was set as β_{1}–β_{2}≠0, where β_{1}–β_{2}>0 indicates a detrimental effect and β_{1}–β_{2}<0 indicates a protection. By using the estimated variance–covariance matrix, the significance of a risk factor was determined using the Wald test statistic.

To build a multivariate model, we used forward selection to evaluate the additive effects of risk factors. The presence of metabolic syndrome, and its individual components, were added one by one into the model. The final model was selected based on the log-likelihood ratio test._{1} and β_{2}, the predicted risk score at time

where _{j}

Between June 2003 and December 2006, 19 812 subjects underwent screening upper endoscopy at NTUH. Of these, 3669 had at least two examinations and comprised our study group (^{2}. GORD symptoms were reported in 11.3%. The prevalence rate (16.4%) of erosive oesophagitis among these 3669 subjects was similar to that (15.7%) obtained from the whole population (n = 19 812) in the previous study.

Characteristic | No. of subjects (%) |

Male | 2483 (67.7) |

Smoker | 416 (11.3) |

Drinks alcohol (at least once per week) | 2219 (60.5) |

Chronic illnesses | |

Cardiac | 47 (1.3) |

Pulmonary | 92 (2.6) |

Hepatic | 501 (13.7) |

Peptic ulcer | 648 (17.7) |

Cholesterol ⩾200 mg/dl | 314 (8.6) |

Hyperuricaemia* and/or history of gout | 283 (7.7) |

Metabolic syndrome | 498 (13.6) |

Enlarged waist circumference | 1239 (33.8) |

Hypertension or blood pressure ⩾130/85 mm Hg | 559 (15.2) |

Diabetes or fasting glucose ⩾110 mg/dl | 198 (5.4) |

HDL-C <40 mg/dl | 1678 (45.7) |

Triglycerides ⩾150 mg/dl | 908 (24.7) |

Exercise, number of times per week | |

⩾5 | 1013 (27.6) |

3–4 | 1585 (43.2) |

⩽2 | 1071 (29.2) |

Sleep quality | |

Good | 1600 (43.6) |

Fair | 1564 (42.6) |

Poor | 505 (13.8) |

Symptoms of GORD | 413 (11.3) |

Short-term use of PPI or H2RA | 587 (11.4)† |

GORD, gastro-oesophageal reflux disease; H2RA, histamine-2 receptor antagonist; HDL-C, high-density lipoprotein cholesterol; PPI, proton pump inhibitor.

*Serum uric acid concentration >7.5 mg/dl.

†From 5145 transition periods.

Subjects underwent up to four endoscopies creating three epochs (baseline endoscopy to endoscopy 2, endoscopy 2 to 3, and endoscopy 3 to 4). The mean duration of epochs 1, 2 and 3 (in days) was 528 (210), 392 (108) and 352 (60).

Epochs | No. of subjects (%) | ||

NE | Class A–B | Class C–D | |

Baseline | Endoscopy 2 (n = 3669) | ||

NE (n = 3066) | 2693 (87.8) | 350 (11.4) | 23 (0.8) |

Class A–B (n = 586) | 249 (42.5) | 304 (51.9) | 33 (5.6) |

Class C–D (n = 17) | 0 (0) | 0 (0) | 17 (100) |

Endoscopy 2 | Endoscopy 3 (n = 1140) | ||

NE (n = 930) | 791 (85.1) | 136 (14.6) | 3 (0.3) |

Class A–B (n = 198) | 74 (37.3) | 109 (55.1) | 15 (7.6) |

Class C–D (n = 12) | 0 (0) | 0 (0) | 12 (100) |

Endoscopy 3 | Endoscopy 4 (n = 336) | ||

NE (n = 252) | 207 (82.1) | 45 (17.9) | 0 (0) |

Class A–B (n = 78) | 27 (34.6) | 39 (50) | 12 (15.4) |

Class C–D (n = 6) | 0 (0) | 0 (0) | 6 (100) |

NE, non-erosive state.

There were 5145 transitions, including 3669 in epoch 1, 1401 in epoch 2, and 336 in epoch 3. Observed rates of transition from NE to erosive oesophagitis were 12.2% (95% CI 8.9% to 15.5%), 14.9% (95% CI 9% to 20.8%) and 17.9% (95% CI 13.2% to 22.6%). The risk of progressing from class A–B to class C–D oesophagitis increased from 5.6% (95% CI 3.7% to 7.5%) in epoch 1 to 15.4% (95% CI 7.4% to 23.4%) in epoch 3; the probability of regression from class A–B to NE decreased from 42.5% (95% CI 38.5% to 46.5%) in epoch 1 to 34.6% (95% CI 24.8% to 44.4%) in epoch 3. However, no statistically significant increase in progression or decrease in regression across epochs was noted (p>0.05). Because no class C–D subjects showed regression to class A–B or NE, the annual regression rate from C–D to A–B (ie, λ_{4}) was set to zero.

Annual progression rates from NE to A–B (ie, λ_{1}) and from A–B to C–D (ie, λ_{3}) were 0.151 (95% CI 0.136 to 0.165) and 0.079 (95% CI 0.063 to 0.094) per person year, respectively, and the regression rate from A–B to NE (ie, λ_{2}) was 0.481 (95% CI 0.425 to 0.536) per person year. The corresponding figures were 0.139 (95% CI 0.126 to 0.152), 0.084 (95% CI 0.066 to 0.101) and 0.346 (95% CI 0.294 to 0.398) when we excluded the transition histories of being administered with short-term PPI or H2RA treatment preceding the next transition.

Model fitting was assessed by comparing predicted with observed transitions using χ^{2}; the lack of a significant difference indicated a good fit for the model (p = 0.415). The observed transition histories were still compatible with the predicted values using the cross-validation method (p = 0.876).

Variables* | RR (95% CI)† | ||

NE→class A–B | Class A–B→NE | Net effect | |

Age ⩾65 years | 1.19 (0.95 to 1.48) | 1.02 (0.78 to 1.34) | 1.17 (0.91 to 1.50) |

Male | 2.36 (1.79 to 3.13)‡ | 0.55 (0.40 to 0.74)‡ | 4.33 (3.30 to 5.66)‡ |

Body mass index ⩾27 kg/m^{2} | 1.28 (1.01 to 1.65)‡ | 0.70 (0.52 to 0.96)‡ | 1.81 (1.36 to 2.41)‡ |

Smoker | 2.27 (1.68 to 3.06)‡ | 1.38 (0.96 to 1.97) | 1.65 (1.24 to 2.18)‡ |

Alcohol use | 1.32 (1.08 to 1.62)‡ | 0.99 (0.78 to 1.26) | 1.34 (1.07 to 1.69)‡ |

Chronic disease | |||

Cardiac | 0.28 (0.07 to 1.16) | 1.19 (0.59 to 2.39) | 0.23 (0.06 to 1.02) |

Pulmonary | 0.98 (0.58 to 1.65) | 0.53 (0.19 to 1.45) | 1.86 (0.72 to 4.80) |

Hepatic | 1.01 (0.76 to 1.34) | 1.09 (0.79 to 1.54) | 0.92 (0.67 to 1.26) |

Peptic ulcer disease | 1.17 (0.92 to 1.49) | 0.88 (0.66 to 1.17) | 1.33 (1.02 to 1.75)‡ |

Cholesterol ⩾200 mg/dl | 0.86 (0.63 to 1.19) | 0.52 (0.34 to 0.79)‡ | 1.66 (1.08 to 2.55)‡ |

Hyperuricaemia and/or history of gout | 1.46 (1.06 to 1.99)‡ | 0.90 (0.61 to 1.34) | 1.61 (1.12 to 2.33)‡ |

Metabolic syndrome | 1.42 (1.11 to 1.80)‡ | 0.76 (0.55 to 0.97)‡ | 1.87 (1.40 to 2.51)‡ |

Enlarged waist circumference | 1.02 (0.84 to 1.25) | 0.78 (0.61 to 0.99)‡ | 1.31 (1.04 to 1.65)‡ |

Hypertension or blood pressure ⩾130/85 mm Hg | 1.32 (1.04 to 1.68)‡ | 0.87 (0.64 to 1.17) | 1.53 (1.15 to 2.04)‡ |

Diabetes or fasting glucose ⩾110 mg/dl | 1.00 (0.68 to 1.48) | 0.68 (0.42 to 1.09) | 1.48 (0.91 to 2.41) |

HDL-C <40 mg/dl | 1.39 (1.15 to 1.68)‡ | 1.00 (0.79 to 1.26) | 1.39 (1.12 to 1.73)‡ |

Triglycerides ⩾150 mg/dl | 1.17 (0.95 to 1.45) | 0.71 (0.55 to 0.91)‡ | 1.66 (1.30 to 2.12)‡ |

Exercise frequency | 0.94 (0.76 to 1.17) | 1.11 (0.87 to 1.45) | 0.84 (0.67 to 1.07) |

Sleep quality | 1.03 (0.77 to 1.38) | 1.12 (0.81 to 1.54) | 0.92 (0.68 to 1.25) |

Symptoms of GORD | 1.23 (0.93 to 1.63) | 0.76 (0.54 to 1.05) | 1.64 (1.18 to 2.27)‡ |

Short-term use of PPI or H2RA | 1.31 (0.86 to 1.97) | 2.83 (2.14 to 3.71)‡ | 0.46 (0.33 to 0.65)‡ |

*Factors were dichotomsed (no/yes) as follows: age ⩾65 years, male, body mass index ⩾27 kg/m^{2}, smoker, alcohol consumed ⩾ once per week, metabolic syndrome, exercise more than twice per week, poor sleep quality, symptoms of GORD and use of short-term PPI or H2RA. The “no” group constitutes the baseline comparator.

†The RR for evaluating the role of each factor was arrived at by taking the exponential of the regression coefficient (β) of the Markov regression—that is, exp(β_{1}) for progression, exp(β_{2}) for regression and exp(β_{1}–β_{2}) for the net effect.

‡p<0.05.

GORD, gastro-oesophageal reflux disease; HDL-C, high-density lipoprotein cholesterol; H2RA, histamine-2 receptor antagonist; PPI, proton pump inhibitor.

Being male (RR = 2.36) and having a BMI ⩾27 kg/m^{2} (RR = 1.28) both raised the likelihood of progressing from NE to erosive disease and lowered the likelihood of regression from erosive to NE disease (RR = 0.55 and 0.70). Smokers and heavy drinkers had a significant risk of erosive disease (RR = 2.27 and 1.32). Subjects with metabolic risk factors, including hypercholesterolaemia, hyperuricaemia, enlarged waist circumference, hypertension, low HDL cholesterol level, hypertriglycaemia and metabolic syndrome, were more likely to progress from NE to erosive disease and/or less likely to regress from erosive to NE states. Short-term PPI or H2RA use increased the likelihood of regression from erosive to NE states (RR = 2.83). GORD symptoms increased the risk of erosive disease (net RR = 1.64).

We used forward selection to evaluate the additive effects of covariates on disease onset and regression. The initial model included only gender (the most significant factor in univariate analysis). We then added variables until they stopped adding significantly to the model. The final model included gender, smoking, metabolic syndrome and short-term PPI or H2RA usage (see

Variables* | RR (95%CI)† | ||

NE→class A–B | Class A–B→NE | Net effect | |

Male | 2.36 (1.73 to 2.97)‡ | 0.53 (0.40 to 0.74)‡ | 4.31 (3.22 to 5.75)‡ |

Smoker | 1.77 (1.32 to 2.36)‡ | 1.48 (0.98 to 2.15) | 1.20 (1.03 to 1.39)‡ |

Metabolic syndrome | 1.29 (1.18 to 1.42)‡ | 0.74 (0.53 to 0.98)‡ | 1.75 (1.29 to 2.38)‡ |

Short-term use of PPI or H2RA | 1.73 (0.92 to 2.77) | 3.19 (2.32 to 4.44)‡ | 0.54 (0.39 to 0.75)‡ |

*Factors were dichotomised (no/yes) as follows: male, smoker, metabolic syndrome and use of short-term use of PPI or H2RA. The “no” group constitutes the baseline comparator.

†The RR for evaluating the role of each factor was arrived at by taking the exponential of the regression coefficient (β) of the Markov regression—that is, exp(β_{1}) for progression, exp(β_{2}) for regression and exp(β_{1}–β_{2}) for the net effect.

‡p <0.05.

H2RA, histamine-2 receptor antagonist; PPI, proton pump inhibitor.

The clinical weight each risk factor contributes to (the net effect of regression coefficients) was 1.46 (natural logarithm of 4.31) for male gender, 0.18 for smoking, 0.56 for metabolic syndrome and −0.61 for short-term PPI or H2RA. The predicted risk score at time

Risk score = (1.46×male)+(0.18×smoking)+(0.56×metabolic syndrome)−(0.61×short-term use of PPI or H2RA)

These dichotomous variables were coded as described in

Kinetic curves can be stratified by classifying predicted risk score into four categories, as shown in

We quantified the natural history of GORD by fitting a large longitudinal follow-up database of patients undergoing endoscopy. The step-by-step transitions are a solid demonstration of GORD’s dynamic nature. The predicted risk score may enable clinicians to develop individually tailored preventive strategies.

In addition to cross-validation, several studies support our model’s credibility on external predictive validity. Among patients with NE reflux disease, 5/33 (15%) developed erosive changes within 6 months

A plausible link can be established between category and continuum theories. The most significant factor affecting vulnerability to erosive oesophagitis is gender. Hence, a slim female who does not smoke or drink alcohol may remain in the NE state for a long time with little chance of developing erosive disease. An obese male, in contrast, who smokes, drinks heavily and has metabolic syndrome (again a typical picture) would probably progress to erosive disease. The probability of changes in disease status being detected at endoscopy would also increase. Thus, different combinations of risk factors lead to different severities of intraoesophageal damage and the disease appears as a continuum upon endoscopic inspection.

The pathogenesis of reflux symptoms is complicated and cannot be explained solely by intraoesophageal damage. Enhanced peripheral and central neural perceptions of stimuli may be crucial in symptom generation.

Obesity significantly increased the risk of GORD symptoms, erosive oesophagitis, Barrett’s oesophagus and oesophageal adenocarcinoma.^{–}

Our results are credible for several reasons. First, we had numerous cases of NE and were able to assess progression to erosive disease. The simultaneous evaluation of symptomatic and asymptomatic subjects also enabled us to observe the entire disease spectrum. Secondly, all our endoscopists completed the same training programme using a standardised rating protocol. This substantially reduced heterogeneity amongst observers and strengthened our ability to model natural history. Thirdly, we found no spontaneous regressions from high-grade erosive states, which conflicts with findings that 42% of patients with class C–D disease regress to A–B and 50% to NE disease within 2 years.

Because progression of GORD is orderly, the Markov approach was appropriate for modelling. However, our target group tended to reflect the general population, so there were few transitions from low- to high-grade oesophagitis and we were unable to investigate the effects of covariates on this stage. Secondly, information about

Our findings suggest that intraoesophageal damage develops as a dynamic process over time. Risk factors in susceptible individuals modulate the likelihood of state-to-state transitions, resulting (upon endoscopy) in the appearance of a continuous spectrum of disease. GORD can therefore be staged with respect to the extent of progression, as with many other chronic diseases. The translation of the quantified knowledge of GORD’s natural history into predicted risk score together with the kinetic curve will be vital for developing individually tailored prevention and screening programmes, to identify candidates for potential interventions, and to determine optimal timing of proposed interventions.

The transition rates in our three-state model can be expressed in an intensity matrix as follows:

where states 1, 2 and 3 represent states NE, LA class A–B and LA class C–D, respectively. The four transition parameters of λ_{1}–λ_{4} have been defined in the text and fig 1. Using the Kolmogorov equation,

where _{ij}_{l}_{l–1}_{1}–λ_{4}, 0 <_{l–1}_{l}

Each individual has his/her own history of endoscopic examination. The likelihood function is established based on the transition probabilities with the use of the data on the history of examinations, in order to estimate the transition parameters λ_{1}–λ_{4}. Such a type of data could be, for example, a man of age 53 years who had undergone repeated endoscopic examinations with the results as follows: NE with smoking on 12 September 2003, class A–B non-smoking on 27 September 2004, NE on 22 July 2005, and NE, again, on 1 June 2006. According to the “lack memory” property of a Markov process, this personal history can be decomposed into three epochs according to the consecutive four endoscopic examinations: (NE→class A–B, 1.04 years| smoking on 12 September), (class A–B→NE, 0.82 years| non-smoking on 27 September), (NE→NE, 0.86 years| still non-smoking on 22 July), as shown for the study group as a whole in table 2. The likelihood of an individual following this history is _{12}(1.04| smoking)× _{21}(0.82| non-smoking)× _{11}(0.86| non-smoking). Note that irregular times are specified for different individuals. To generalise, if _{0}, _{1}, _{2} and _{3}, creating three epochs (_{1}–_{0}), (_{2}–_{1}) and (_{3}–_{2}), the total likelihood function for all individuals can be written as:

where the _{ijl}_{l}_{–1} and in state _{l}. The maximum likelihood estimates of the transition rates can be obtained using the Newton–Raphson procedure.

Given the estimated annual progression rates from NE to class A–B oesophagitis and from A–B to class C–D oesophagitis of 0.151 (λ_{1}) and 0.079 (λ_{3}) per person year, respectively, and the regression rate from class A–B to NE and from C–D to class A–B of 0.481 (λ_{2}) per person year and 0 (λ_{4}), respectively, the transition rate matrix can be expressed as follows:

The cumulative risk (probability) over time can be calculated using an SAS/IML procedure.

where the upper and middle rows indicate the first year cumulative risks depicted in the kinetic curve 2A and 2B, respectively.