On the possibilities to reduce automatic alerts during electronic prescription
Wolfgang B. Lindemann, Cabinet de médecine générale, 67 rue du Maréchal Foch, F – 67113 Blaesheim
One task of electronic health records (EHR) is decision support during electronic prescription by generating
automatic alerts triggered by preexisting diseases, allergies or drug-drug-interactions. Those alerts are generally
too abundant and too unspecific to be usefull1 2 3. It has been evaluated in a university hospital setting2 3 which
information already present in the EHR could improve their quality. I repeated this study in general medicine.
Table 1
Alert avoidable by
consideration of
Number Example(s) of avoidable alert(s)
(%) of
all alerts
renal elimination
capacity
198
(12.1%)
In the elderly, the MDRD or the CKD-EPI –
formulas give a better estimation of the renal
clearance: the Cockcroft-Gault-formula has a
strong tendency to underestimate the clearance.
Furthermore, there is mostly no clear cut-off point
of 90, 60 or 30 ml/min contraindicating a given
drug as assumed by the VIDAL database, but the
individual patient profile must be considered.
# Axisanté/Vidal base
the alerts exclusively
on the CockcroftGault clearance and
without further
considering the
specific patient profile.
unclear reason
138
(8.40%)
First prescription of insulin for a drug-naïve type-Idiabetic displays “contraindication because of
association with pioglitazone”. An ACE inhibitor
displays “contraindicated in arterial hypertension”.
# I did not or not
fully understand the
reason for the alert
itself or its trigger.
protective
comedication
59
(3.59%)
NSA with proton pump inhibitor.
Antiepileptic treatment already present in a patient
receiving an epilepsy-threshold lowering treatment.
#
chronology of and
lag time between
administrations
52
(3.17%)
Levothyroxine, bisphosphonate or topic gastric
antacid taken at distance from other medications
#
dose
45
(2.74%)
Acetylsalicylic acid in cardiac dose (75-160mg).
Nitrate spray with benzodiazepine because the
spray contains traces of alcohol.
#
last blood pressures
43
(2.62%)
Bisoprolol and tamsulosin in a patient with stable
blood pressure (and receiving amlodipine for
hypertension). Long-time coprescription of
ramipril and indapamid.
#+
age
36
(2.19%)
Statins, sartans or anticoagulation display
“contraindicated/not recommended in fertile
women”. Zolmitriptan displays “caution in women
after menopause”.
##
course of therapy
33
(2.00%)
Risk of hypotension only after instauration/change
of a combination therapy diuretic plus sartan
##
last HbA1c values
21
(1.28%)
Hydrochlorothiazide or ACE-inhibitor in an
equilibrated diabetes
## +
administration route 14
(0.85%)
Cutaneous, bronchial or other topic application
with little systemic passage
##
last heart
frequencies
12
(0.11%)
Atenolol in patient with incomplete right bundle
branch block
## +
last fasting blood
sugars
13
(0.79)
Hydrochlorothiazide or ACE-inhibitor in an
equilibrated diabetes
## +
last TSH values
2
(0.12%
Alert “evaluation of under- or overdose of levothyroxine does not consider adaptation to the elderly”
## +
last INR values
2
(0.12%)
Alert for higher than normally recommended dose
of warfarin in a patient needing this dose for INR
## +
(respective
prescription: figure 1)
Material and Methods
For 86 randomly selected patients who choose me as family doctor I evaluated how many automatic alerts for
their long-term-medication would be avoidable if information usually present in the EHR would have been
considered. I further evaluated the alerts of up to day 27 randomly selected intercurrent medications (data not
shown but similar). I had seen all patients at least 5 times during at least 3 months (the time necessary to
complete the EHR in structured form). The statistical evaluation was done with Excel 2016 and SPSS 231.
Results
The 86 patients (43 women) were aged 17-88, mean 60. They had a mean number of 4.2 structured preexisting
conditions and received a mean number of 5.6 medicaments as long-term medication. Older patients had
more structured preexisting conditions** r=0.44 and received more medicaments** r=0.46. (I have not gathered all
preexisting conditions in structured form but it is unlikely that there is any bias in the selection of patients
whose pathologies are already structured; in any case, at least 85% of them should have been gathered).
Patients having more structured preexisting conditions received more medicaments** r=0.60.
480 prescribed medicaments generated in the prescription window (figure 1) a total number of alerts of 342 of
low, 397 of middle and 535 of high alerting level: a mean of 0.71, 0.83, 1.12 alerts per prescription
respectively, letting 37.9%, 35.2% and 26.9% prescriptions respectively alert-free at low, middle and high level.
During the conduct of the study (december to august 2016, the first long-term medication which was saved
for evaluation with all its alerts was generated on 1st of February 2016), there were 1 to 2 monthly updates of
the underlying Vidal database and the number of low* r=0.30- and high-level* r=0.28 alerts decreased indicating
probably an ongoing improvement of the database.
A total of 2170 alerts showed up in the “Vidal résumé”-window (a detailed list of all alerts for a given
prescription with explication of the cause). 527 of them were irrelevant because the medication triggering the
alert was not on the actual prescription. Of the remaining 1643 alerts, 214 (13.0%) were doubles# and 245
(14.9%) “insufficient data for checking the prescription” (indicating a problem of the database and not of the
prescription): 182 (11.1%) just mentioned “database insufficient for evaluation”, in the remaining 63 cases the
system most frequently stated “database insufficient for evaluation in the elderly”. During the study, the
number of alerts for “insufficient data” decreased significantly** r=-0.52 probably indicating an ongoing
improvement of the database. 18 (1.11%) alerts appeared because the EHR did not permit a structured entry
of the posology (for example “once a week”) ##, 14 (0.85%) alerts were not justified because of an on-demand
medication##. Finally, 97 (5.90%) alerts showed up because the system misinterpreted a structured pathology#,
further indicating a problem of the database used (for example identifying (treated) “hemochromatosis” with
“liver cirrhosis/insufficiency” or “arterial hypertension” with “hypertensive cardiomyopathy”).
I judged 357 (21.7%) alerts as unavoidable, most of them were related to the age of the patient as “special
monitoring in the elderly necessary” and in 3 cases (0.18% of all 1643 relevant or 0.14% of all 2170 alerts in
the “Vidal-résumé window) I changed a treatment according to a justified alert which I had overlooked during
routine care because it had been buried in the wealth of useless or wrong alerts (see table 1). So 3 of my 480
prescriptions (0.63%) were improved by the automatic alerts, plus maybe a dozen prescriptions which I avoided during prescribing because the prescription window indicated “absolute contraindication”. No patient
had suffered any harm. During the conduct of the study I did not consider less alerts as unavoidable p<0.88 r=0.017.
For 668 of the remaining alerts (40.7% of all relevant 1643 alerts) I evaluated how information already present
in the EHR could have avoided them (table 1). Only the number of posology alerts decreased significantly
during the study** r=0.54 suggesting ongoing improvement of the database (I did not change prescribing habits).
docteur@blaesheim-medical.fr
Comments
1 www.spss.com * p<0.05 ** p< 0.01 two-sided Pearson
Discussion
I found no similar study in general medicine4, Seidling (2014) and Czock (2015) seem to be the only such studies
in a hospital setting. The automatic alerts are created in interaction between my EHR Axisanté 5 (Compugroup)
and the common French database “VIDAL”, comparable to the “Rote Liste” in Germany. Since 2012, the 11th
convention incites French general practitioners to use an EHR like the one I evaluated generating automatic
alerts during prescription. The alert quality is not only bad but ridiculous (as reported in many former studies and
this “alert overkill” may, on the top, still miss serious medication errors5 6) so that this incitation may not lead to
improvement of medical practice7 but only to a greater cognitive workload8. This study supports my hypothesis
in a former study9 10 that the rather low self-reported reactivity on automatic alerts was still too high to be true.
Main limitation of my study is that I am no trained pharmacologist and it should be externally evaluated whether
really only 3 of 480 of my prescriptions needed improvement (but most of the patients with more severe
conditions are seen at least once a year by a specialist who will supervise the long-term medication).
Nevertheless, the evaluation of a medicament’s benefit-risk-profile is part of general practice. I have only
formalized and applied those criteria we are using in practice and whose general security is self-checked during
the follow-up of the patient. I would expect a pharmacologist to find more possibilities to filter unnecessary
alerts because in the doubt I considered an alert as unavoidable, which is reflected by their still relatively high
number. There is both a high need and a high potential for improvement of automatic medication alerts.
.
Total:
668 (40.7%)
+ considered as safe if the patient had a stable normal value (creating an
algorithm determining if the value is stable and normal with respect to the patient profile will be challenging)
# did not change during the conduct of the study ## change during the conduct of the study not evaluated
References
1 Sittig DF, Longhurst CA et al. Electronic health record features, functions and privileges that clinicians need to provide safe
and effective care for adults. In: Weaver CA, Ball MJ, Kim GR, Kiel JM (editors). Healthcare Information Managament
Systems: Cases, Strategies, and Solutions, Springer International 4th edition 2015
2 Seidling HM, Klein U, Schaier M, Czock et al. What, if all alerts were specific – Estimating the potential impact on alert
burden, Int J Med Inf 2014; 83: 285-291
3 Czock D, Konias M, Seidling HM et al. Tailoring of alerts substantially reduces the alert burden in computerized clinical
decision support for drugs that should be avoided in patients with renal disease, J Am Med Inform Assoc 2015; 22:881-887
4 My search strategy was as follows: First, I conducted a Pubmed-search with the logic “(EHR or (electronic health record)
or (electronic health records) or (electronic patient record) or EPR or (clinical decision support system) or (clinical
decision support systems) or (health information technology)) AND ((primary care) or (ambulatory care) or (family
medicine) or (general practitioner)) AND (alert or alerts or (alerting function) or (alerting functions)) AND ((alert
specificity) or (alert fatigue) or (alert reduction) or reduction or specificity).Then, I searched in Pubmed, researchgate.net
and the respective journal websites articles citing the articles of Seidling (2014) and Czock (2015), in Pubmed I also
searched for similar articles. Finally, I searched manually the years 2010-2016 in the 5 journals Applied Clinical Informatics,
Methods of Information in Medicine, Journal of the American Medical Informatics Association, International Journal of
Medical Informatics, Journal of Innovation in Health Informatics (formerly: Informatics in Primary Care)
5 Slight S, Eguale T, Amato MG et al. The vulnerabilities of computerized physician order entry systems: a qualitative study.
J Am Med Inf Ass 2016; 23 (2): 311-316
6 Topaz M, Segler DL, Slight SP et al. Rising drug allergy overrides in electronic health records: an observational
retrospective study of a decade of experience. J Am Med Inf Ass 2016; 23 (3): 601-608
7 Bayoumi I, Balas MA, Handler SM et al. The effectiveness of computerized drug-lab alerts: A systematic review and metaanalysis, Int J Med Inf 2014; 83 (8): 406-415
8 Ariza F, Kalra D, Potts HWW. How do clinical information systems affect the cognitive demands of general practitioners?
Usability study with a focus on cognitive workload. J Innov Health Inform. 2015; 22(4): 379-390.
9 Lindemann WB Über (Aus)nutzung von Arztinformationssystemen und ärztliche (Lebens)qualität). Poster 83: 60th Annual
meeting of the GMDS 9/ 9/ 2015 (Krefeld/Germany). (On the use of electronic health records and (life) quality of primary
care physicians). German non English. Lindemann WB Über (Aus)nutzung von Arztinformationssystemen in der hausärztlichen Praxis. Poster 201: 59th Annual meeting of the GMDS 9/ 8/ 2014 (Göttingen/Germany). (On the use of electronic
health records in primary care). German non English.
Both available on www.researchgate.net “Wolfgang B. Lindemann”.
Figure 1: A prescription window at middle-level alerting (as I considered “niveau 4” in this study)
absolute contraindication
relative contraindication
posology alert
overdose alert
interaction alert
no alert
10 Nanji KC, Slight SP, Seger DL et al. Overrides of medication related clinical decision support alerts in outpatients. J Am
Med Inform Assoc 2014; 21: 487-481
Contribution 368 presented on 8/29/2016 at HEC 2016 in Munich. This poster, its presentation
and future journal publications are available on www.researchgate.net ”Wolfgang B. Lindemann“.