Can routine information from electronic patient records predict a future diagnosis of alcohol use disorder?

OBJECTIVE: To explore whether information regarding potentially alcohol-related health incidents recorded in electronic patient records might aid in earlier identification of alcohol use disorders.
DESIGN: We extracted potentially alcohol-related information in electronic patient records and tested if alcohol-related diagnoses, prescriptions of codeine, tramadol, ethylmorphine, and benzodiazepines; elevated levels of gamma-glutamyl-transferase (GGT), and mean cell volume (MCV); and new sick leave certificates predicted specific alcohol use disorder.
SETTING: Nine general practitioner surgeries with varying size and stability.
SUBJECTS: Totally 20,764 patients with active electronic patient record until data gathering and with a history of at least four years without a specific alcohol use disorder after turning 18 years of age.
METHODS: The Cox proportional hazard analysis with time-dependent covariates of potential accumulated risks over the previous four years. MAIN OUTCOME MEASURES: Time from inclusion until the first specific alcohol use disorder, defined by either an alcohol specific diagnostic code or a text fragment documenting an alcohol problem.
RESULTS: In the unadjusted and adjusted Cox-regression with time-dependent covariates all variables were highly significant with adjusted hazard ratios ranging from 1.25 to 3.50. Addictive drugs, sick leaves, GGT, MCV and International Classification for Primary Care version 2 (ICPC-2), and International Classification of Diseases version 10 (ICD-10) diagnoses were analyzed. Elevated GGT and MCV, ICD-10-diagnoses, and gender demonstrated the highest hazard ratios.
CONCLUSIONS: Many frequent health problems are potential predictors of an increased risk or vulnerability for alcohol use disorders. However, due to the modest hazard ratios, we were unable to establish a clinically useful tool. KEY POINTS Alcohol is potentially relevant for many health problems, but current strategies for identification and intervention in primary health care have not been successful. Many frequent clinical problems recorded in electronic patient records may indicate an increased risk for alcohol related health problems. The hazard ratios were modest and the resulting predictive model was unsatisfactory for diagnostic purposes. If we accepted a sensitivity as low as 0.50, the specificity slightly exceeded 0.75. With a low prevalent condition, it is obvious that the false positive problem will be vast. In addition to responding to elevated blood levels of liver enzymes, general practitioners should be aware of alcohol as a potentially relevant factor for patients with repeated events of many mental and psychosocial diagnoses and new sick leaves and repeated prescriptions of addictive drugs.

Introduction

General practitioners (GPs) as health care providers for the general public are important actors in dealing with alcohol-related health problems.[1] The link between alcohol consumption and numerous health problems is strong, and earlier identification of risky or harmful drinking is regarded essential, both in public health terms and for the individual patient.[2-4] The preferred method for identification and treatment of risky or harmful drinking has, for the past decades, been screening and brief intervention (SBI), but important questions concerning the effectiveness of SBI in routine health care settings remain unanswered.[5,6] Furthermore, recent large scale implementation studies have failed to show effect.[7-9]

However, it is known that GPs regard dealing with alcohol-related health problems a legitimate part of their responsibility.[10-13] The recent recognition of the lack of robust evidence for SBI in routine health care settings necessitates further research. Recently, approaches based on clinical relevance instead of screening measures have been studied.[14-16] Health incidents or changes in the patient’s life are used as indicators of potential relevance for addressing alcohol. These have been coined as pragmatic case finding [14] or semi-systematic method,[15] they not only focus primarily on clinical signs, but also focus on targeted screening in some routine situations. These strategies focus on the present clinical situation and the awareness that alcohol may be relevant for a patient’s health, both as a possible cause and as a complicating factor for their health problems. GPs struggle with asking about alcohol out of context, as in general screening, but asking based on potential relevance in a specific clinical situation is probably a better foundation for interventions.[17,18]

In general practice, the patient records will often contain information gathered through many years. Almost all GPs in Norway use electronic patient records (EPR), but the systems are not highly functional in systematizing relevant information.[19] Not only the patient’s present health problem, but also previous incidents may indicate relevance for talking about alcohol.

The aim of this study was to explore whether information regarding potentially alcohol-related health incidents recorded in EPRs might aid in earlier recognition of alcohol related health problems.

Design, methods, and material

Nine GP surgeries in the Stavanger region in south west of Norway were recruited. They were chosen on the basis of maximal variety in size (1–7 doctors) and stability (high turnover to high stability), and all had applied EPR for at least 10 years. The total number of doctors was 36. All 20,764 patients with an active EPR (alive and registered with a doctor the month prior to data gathering) were included. Gender, year of birth, name of registered doctor and surgery were registered, and the patients were given a unique, non-reversible eight-digit code with letters and numbers. After completion of data gathering, the eight-digit code was replaced and the patients were consecutively numbered. Doctors and surgeries were also numbered.

A vast majority of GP surgeries in the region at that time used the same EPR system, and a computer program designed to extract data from records in this system was made. We tested a pilot version of the computer program in the largest surgery during March 2011, and then an automatized version was tested and applied a few months later.

We extracted data that might be alcohol-related to test these data against comprehensive alcohol use disorder (c-AUD) as endpoint. C-AUD was defined as either an alcohol use disorder (AUD) according to ICPC-2 or ICD-10 [3,4] or a text fragment (AUD text fragment) documenting that an alcohol problem was dealt with. In Norway, ICPC-2 is applied in general practice, whereas the specialized health care system applies ICD-10. ICD-10-diagnoses for AUD were translated to ICPC-2 in 2010 applying standardized tables,[20] in order to identify ICPC-2-diagnoses for AUD. The diagnostic codes in ICD-10 are more specific (three to four figures) compared to ICPC-2 (two figures), thus to retain the specificity of AUD in ICD-10 we used the corresponding ICPC-2-code solely if “alcohol” was included in the text-field of the ICPC-2-diagnosis. See Appendix A for all diagnoses included in AUD.

Clinical experience indicates that the threshold for identifying an AUD with a formal diagnosis may be high in general practice. We wanted to include as outcome situations where an AUD was documented in the running text of the EPR, but where no formal diagnosis was made. We identified the word “alcohol” in the running text, either alone or as a compound word (in Norwegian compound words are frequently used when the English expression would contain two or three words, e.g., “alkoholmisbruk”, English: alcohol abuse). Compound words highly indicative of an AUD were defined as an AUD text fragment. All versions of compound words containing “alcohol” were assessed manually and either defined as an AUD text fragment or not. The validity of this AUD text fragment was tested by performing a second data collection in one surgery to explore the context of the AUD text fragment by manually assessing a 12 word text fragment with the compound word with “alcohol” in the middle. This was done for a three-year period (January 2001 to December 2003) in one of the surgeries of medium size. We found 102 fragments which had been defined as AUD text fragments, and for 20 of these (20%) it was evident that the alcohol problem in question was someone else’s, most frequently a parent. We also identified 171 text fragments with “alcohol” originally not identified as an AUD text fragment. Of these, as many as 105 (60%) dealt with a real alcohol problem for the patient. Many of the patients had several such text fragments. This suggests that our method of defining an AUD text fragment is more prone to underestimate than overestimate the prevalence of an AUD.

The term c-AUD was defined as either an AUD or an AUD text fragment or both and used as outcome for the analyses. Censoring date was defined as the first of the month prior to data gathering, or the last predictor event if more recent. Start of follow-up (t = 0) for all patients were defined after an observation period of four years free from c-AUD in the record.

Predictors were firstly potentially alcohol-related ICD-10-diagnoses with attributable fractions larger than 0.3.[3,4] These diagnoses were translated to ICPC-2, with a consequently lower precision level due to the wider categories of ICPC-2. We included other ICPC-2-diagnoses where there is evidence of a potential causal relation with alcohol consumption.[15,21,22] See Appendix A for all diagnoses used as predictor events. Other predictors were number of new sick leaves, non-narcotic controlled substances (class B-drugs in Norway) and elevated blood levels for GGT and MCV.[23-26] A new sick leave was defined as a full time (not partial) sick leave with at least 16 days since a previous sick leave. Class B-drugs were the non-narcotic controlled substances codeine, tramadol, ethylmorphine, and benzodiazepines, including z-drugs. Gender was included as predictor.

All patients had a total history of 4–21 years, and all had an active patient record until data collection. For patients with a record prior to the age of 18, their observation period started from 1 January, the year they turned 18. Observations stopped at the age of 80 years. All readable data in the EPRs were scanned by the program, including incoming reports.

Statistical methods

For descriptive statistics, we used mean, median, and range. Correlations were estimated by Spearman’s rho.[27] Time from inclusion to c-AUD was analyzed applying the Cox proportional hazards model [28] including time-dependent covariates.[29,30] The covariate values were updated at each time point for the following types of predictor events: B-drugs, new sick leaves, elevated blood levels for GGT and MCV, alcohol-related ICPC-2 and ICD-10 diagnoses. Thus, the following variables were included in the Cox-regression models: gender, number of new sick leaves, number of prescriptions of class B drugs, number of elevated GGT and MCV levels, number of alcohol-related ICD-10 diagnoses and number of alcohol-related ICPC-2 diagnoses.

To do the analyses, the data file was organized in long format with one line per event date and varying number of lines per patient. Both simple and multiple Cox-regression were run. Results are reported as unadjusted and adjusted hazard ratios (HR), respectively, with 95% confidence intervals (CI) and p values from Wald tests. The analyses were done using Stata 13 (College Station, TX) and all predictors were reported per 10 predictor events. We excluded from the model predictor events more than four years prior to the present predictor event.[31] Predictor events prior to t = 0 were summed up and added to the events, however they were also gradually excluded during the first four years after t = 0.

From the final multiple Cox-regression model, a prognostic index was defined equal to the fitter linear predictor equation in the model. Receiver operator characteristics (ROC) of this index was evaluated against the patients’ c-AUD status four years after each update of the index (i.e., new predictor event) by calculating sensitivity and specificity and plotting the corresponding ROC curve.[27]

Results

The 20,764 patients, 43% of which were males, had follow-up times of up to 17.0 years after t = 0, with a median of 12.5 years (Table 1). The maximum number of events for each predictor is very high, though the medians are low, demonstrating that most patients have a small number of events for each predictor. 2.9% of the patients had a positive end point (c-AUD), of which 53.3% male. When splitting up, we found that 43% of these had an AUD (1.3% of all patients, 67.9% male), whereas 57% had only AUD text fragment (1.6% of all patients, 41.1% male).

Table 1.

Descriptive statistics for n = 20,764 patients from nine general practice surgeries in the Stavanger area in Norway accrued from March to August of 2011.

Variables	Mean	Median	Range
Follow-up time, years	6.5	12.5	0.0–17.0
Age at start of follow-up, years	43.4	42.0	22–79
Born	1956	1957	1916–1988
No. of predictor events
Class B drugs	28.6	5	0–774
New sick leave	5.9	3	0–143
Elevated laboratory test	0.7	0	0–66
ICD-10-diagnoses	0.3	0	0–50
ICPC-2-diagnoses	2.4	1	0–130
Cumulative predictor events	37.8	16	0–870

Abbreviations: ICD-10: International Classification of Diseases, version 10; ICPC-2: International Classification of Primary Care, version 2.

In the simple Cox-regression, all variables were significant, and only class B prescriptions and gender had an HR lower than 2 per 10 events (Table 2). In the adjusted Cox-regression the HR was highest for elevated blood tests for GGT and MCV with 3.5 per 10 events, and just below 2 for ICD-10 diagnoses, gender, and new sick leaves. All variables were highly significant. The lowest estimates were class B drugs and ICPC-2 diagnoses.

Table 2.

Results from Cox regression of alcohol use disorder with time-dependent covariates for 20,764 patients from nine general practice surgeries in the Stavanger area in Norway accrued from March to August of 2011.

	Unadjusted estimates			Adjusted estimates
Variables	HR	95% CI	p	HR	95% CI	p
Gender (male)	1.71	(1.46, 2.01)	<0.001	1.94	(1.65, 2.29)	<0.001
No. of predictor events
Class B drugsa	1.27	(1.24, 1.31)	<0.001	1.25	(1.21, 1.28)	<0.001
New sick leavea	2.16	(1.81, 2.58)	<0.001	1.81	(1.50, 2.19)	<0.001
Elevated lab testa	3.62	(2.93, 4.46)	<0.001	3.50	(2.79, 4.39)	<0.001
ICD-10-diagnosesa	2.51	(1.51, 4.18)	<0.029	2.00	(1.07, 3.72)	0.009
ICPC-2-diagnosesa	2.29	(2.01, 2.61)	0.001	1.43	(1.16, 1.78)	0.002

Abbreviations: HR: hazard ratio; CI: confidence interval; ICD-10: International Classification of Diseases, version 10; ICPC-2: International Classification of Primary Care, version 2.

Per 10 predictor events.

We made a prognostic index from all significant regression coefficients in the adjusted model. ROC of this index compared to status four years later gave a fairly modest area under the curve (AUC) of 0.72 (Figure 1).[32,33] The curve reveals that with a sensitivity of 0.5, the corresponding specificity is slightly above 0.75.

Figure 1.

Receiver operator characteristics (ROC) curve for prognostic index (gender, elevated lab tests, class B-drugs, new sick leaves, and alcohol-related ICPC-2 and ICD-10 diagnoses), for n = 16,814 patients from the Stavanger area in Norway, for comprehensive alcohol use disorder. Abbreviations: ICD-10: International Classification of Diseases, version 10; ICPC-2: International Classification of Primary Care, version 2.

Discussion

Our findings show that repeated incidents of many common clinical problems in general practice represent an increased risk of identifying an AUD later on, but the results are not strong enough to enable the development of a clinically relevant identification strategy. Elevated blood tests of GGT and MCV, new sick leaves, prescriptions of class B drugs, and a wide variety of diagnoses were significantly associated with increased risk of a future AUD, though the HRs were fairly modest. All predictors represent frequent incidents in general practice, where the patient trajectories often are long.

Sample and methods

In this study, we have included all patients who had an active EPR until data collection and for at least four years after they turned 18 years of age instead of collecting data on a sample of eligible patients. This ensures realistic data. The variables were chosen with adults in mind, and data prior to the year they turned 18 was therefore not included in the material. In old age, the number of health problems rapidly increase, and collecting data for the prediction of future health problems is less relevant.

Other data from the EPRs than the chosen variables were not collected, apart from gender, year of birth, doctor and surgery, and first and last entry in the EPR. We chose to exclude predictor events more than four years prior to a predictor event, because recent events in a patient’s life, documented in the EPR, probably have a higher impact on present health. This view was supported by the fact that the HRs were lower when performing Cox-regressions without exclusion of predictor events more than four years prior (analyses not shown).

Significance of the results

We wanted to explore whether clinical information, as recorded in the EPR, might aid the doctors in establishing relevance for addressing alcohol. Several HRs were around 2, though the analyses were done per 10 predictor events. But the number of such events for a patient may be very high, and all predictors represent frequent clinical problems in general practice. Many different kinds of events sum up the risk as the model indicates. Our validation of the AUD text fragment indicated that our definition of c-AUD is underestimating the diagnosis.

Gender was more strongly associated with AUD than with c-AUD, indicating a lower threshold for applying a specific diagnosis to a recognized alcohol problem if the patient is male. A gender difference in SBI is also described in a Cochrane review, but whether the gender difference is primarily caused by identification or treatment differences is not known.[5,34] Elevated blood levels of GGT and MCV is a recognized starting point for alcohol talks in general practice, and their relevance has been tested in previous studies.[14,24,35] Such changes are late effects of too high alcohol consumption, and many psychosocial problems may occur much earlier.[15,36] We found low estimate for ICPC-2 diagnoses, perhaps because this is a composite variable, composed by converted diagnoses from ICD-10,[3,4] pragmatic case finding,[14] and early clinical signs [15] (Appendix A).

The AUC of the ROC-curve was fairly modest, and the direct clinical relevance is modest. We have chosen to exclude predictor events more than four years prior to the present event. Previous events sum up and constitute an ever increasing risk, but previous difficulties and problems are also overcome and sometimes balanced by positive experiences. While events early in life may have strong effects on present and future health, information in the EPR will probably not be a strong indicator of relevant events in early lifetime.[37]

If our choice of predictors has been adequate, the results indicate that using patient record data to establish a threshold value for identifying an AUD is futile because of lack of sensitivity and specificity. But our findings point to the fact that many frequent clinical problems normally not conceived as caused by alcohol consumption, over time may be related to alcohol consumption. The predictor events constituting a potential risk, as well as the opportunities to intervene, increase over time in general practice. Even interventions with minor effect may potentially add up in the long run, when applied many times and for many patients.[38,39]

SBI has shown a lack of diagnostic accuracy, intervention efficacy, and feasibility.[8,11,17,40] Methods focusing on the present situation and the patient’s problem will probably increase relevance and recognition for the patient.[14,15,18] We should bear in mind that alcohol use may represent attempts to master a challenging life as viewed from the patients’ perspective.[41] An open and respectful dialogue is needed to explore how alcohol may be relevant for health, coping and well-being.

Strengths and limitations of the study

The large variety in size and stability for the surgeries supports external validity.[42] The extensive number of 36 participating GPs and 20,764 patients together with the long observation period of 4–21 years further strengthens the external validity.

Being an exploratory study in EPRs, the data is highly affected by everyday habits, flaws and inaccuracies in diagnostic work, interventions and documentation. Because of a maximum observation period of 21 years, many doctors have been replaced over the years, thus several doctors may have been responsible for each patient’s EPR. The resulting diagnostic variability probably reduces the internal validity of the study, but strengthens the external validity.[42] When facing uncertainty, the result of the diagnostic process will vary greatly between doctors.[43]

Many address alcohol and document the interventions without proper diagnosis, but it is also likely that many interventions are not documented in the EPR. In addition, we have also demonstrated that c-AUD underestimated real alcohol problems in patients at least in the EPRs that were examined. Since we have no direct assessments of the patients, we know nothing about the real prevalence of alcohol related health problems in the material. As the data are completely anonymous, we are not able to test our data against registers as the Cause of Death Registry or the Norwegian Registry Database, nor may we trace a patient moving from one surgery to another.

Sick leaves were difficult to trace uniformly by the data extraction software because of the extensive time frame, and many sick leaves, especially before 2000, had to be excluded. The results for sick leaves are therefore less robust. The composite variable of ICPC-2 diagnoses obscures the potential relation between diagnoses or clusters of diagnoses and AUD. In future studies, ICPC-2-diagnoses should probably be grouped in clinically meaningful clusters in order to be able to detect stronger correlations than we have found. The screening of text fragments may also be more extensively utilized in further studies, to explore relations between alcohol-related health problems and different clinical situations described but not diagnoses in the EPR.

Implications

We have shown that many everyday health problems may, over time, indicate an increased risk of a future AUD. The variables explored in this study may be just as important as vulnerability factors as they are potentially early signs of an alcohol-related health problem. Our findings emphasize the importance of asking about alcohol consumption in many common clinical situations, exemplified by the variables in this study. The unsatisfactory diagnostic accuracy precludes the development of a clinically useful tool, but this is not a valid objection to asking about alcohol consumption based on potential relevance.

Many patients may be aware of the possible relation between their health problem and alcohol consumption.[38] Other patients may be unaware of such a relation. A GP addressing this possible relation in an open, non-judgmental manner may represent one of many important elements in a long and winding road to permanent change.