There were a total of 1,753 datapoints from 756 patients that had simultaneous determination of PHTtotal and PHTfree and also a plasma albumin measured within 7 days of the PHT measurements. As shown in Table , the patient population studied was mostly 15 years or older (n = 701 out of 756). Roughly equal numbers of patients were being administered monotherapy with phenytoin for seizure control (n = 386) as compared to being prescribed one or more additional anti-epileptic drugs in addition to phenytoin (n = 370). The most common co-administered anti-epileptic drugs were levetiracetam (167 patients), phenobarbital (51 patients), and valproic acid (44 patients) (Table ). At the time of blood draw for the initial phenytoin drug level, 263 patients had documentation of seizure within 24 hrs while 493 patients did not.
Figures and show scatterplots of PHTfree
,, respectively, using only the initial laboratory data for patients (i.e., not including repeat measurements for patients). The Pearson coefficient for the correlations between PHTtotal/10
was 0.72 and for PHTfree
was 0.79. The slope of the regression line for the relationship between PHTfree
was statistically different than the line of identity (slope = 1) (95% confidence interval: 1.045-1.163, P < 0.05 for comparison of slope to 1). In contrast, the slope of the regression line for the relationship between PHTfree
was not statistically different from 1 (95% confidence interval: 0.926-1.034). Additional file 1
: Figure S1 (found in Additional file 1
) shows plots similar to Figure and except using all laboratory data, including repeated measurements (i.e., using all 1,753 datapoints from 756 patients). These present very similar relationships to that seen in Figure and . Additional file 1
: Figure S2 presents Bland-Altman (difference) plots of the data in Figure , both by absolute (Additional file 1
: Figure S2A) and percent bias (Additional file 1
: Figure S2B).
Figure 1 Correlation of total, adjusted, and free phenytoin plasma concentrations. (A) Correlation of PHTfree versus PHTtotal/10 for 756 patients is shown. For patients who had multiple phenytoin measurements, only the chronologically first set of PHTfree and (more ...)
Although the PHTadj_free
provides a higher Pearson correlation to PHTfree
, the question is whether a clinician is more likely to make a different decision when presented with PHTadj_free
as opposed to PHTtotal
. This was investigated through the use of contingency tables. Three-by-three contingency tables were constructed comparing grouping of results for PHTtotal
with respect to their therapeutic ranges (10-20 mg/L for total phenytoin; 1-2 mg/L for free phenytoin). As Figure and Additional file 1
: Figure S3A shows, PHTtotal
frequently is in a lower category than PHTfree
below its therapeutic range but PHTfree
within or above its reference range), which could lead to clinical overdosing of the patient if PHTtotal
and not PHTfree
were used as the basis to guide dosing. The converse situation (PHTfree
in a lower category than PHTtotal
) was uncommon. Overall, PHTfree
were concordant with respect to therapeutic category less than 50% of the time (43.1% for all datapoints and 46.6% when excluding repeated measurements).
Figure 2 Clinical decision using total phenytoin or adjusted phenytoin compared to measured free phenytoin. (A) The data is derived from three-by-three contingency tables comparing grouping of PHTtotal and PHTfree into lower than therapeutic reference range (L), (more ...)
On the other hand, three-by-three contingency tables showed that PHTadj_free
had improved concordance, relative to PHTtotal
, to PHTfree
with respect to therapeutic category (Figure , Additional file 1
: Figure S3B). Similar to the analysis between PHTtotal
was more frequently in a lower category to PHTfree
than in a higher category. However, PHTfree
were concordant nearly 70% of the time (68.7% for all datapoints and 69.6% when excluding repeated measurements), statistically superior to PHTtotal
(Fisher's exact test < 0.001).
The concordance data was also broken down into patients who did or did not have documented seizures within 24 hours of the blood draw for phenytoin drug level (Additional file 1
: Figure S4) and those on phenytoin monotherapy for seizure therapy versus those also being treated with additional anti-epileptic drugs (Additional file 1
: Figure S5). PHTfree
were concordant 73.2% for patient without recent seizures but only 56.6% for those who had seizures within 24 hours (Fisher's exact test < 0.001). In contrast, PHTfree
were concordant 69.9% for patients on phenytoin monotherapy and 69.3% for patients on polytherapy for seizure control (Fisher's exact test > 0.05).
We also looked at the three-by-three contingency table data to see how stable phenytoin measurements were for patients who had multiple phenytoin measurements over time. In particular, we compared how often, for a given patient, the temporally next phenytoin measurement fell in the same category in the three-by-three table as the previous set of measurements. For the data comparing PHTtotal with PHTfree, the next consecutive set of phenytoin measurements agreed 53.7% of the time with the previous measurements (529 out of 997 measurements). For the data comparing PHTadj_free with PHTfree, the next consecutive set of phenytoin measurements agreed 46.9% of the time with the previous measurements (468 out of 997 measurements). These data may reflect the predominantly inpatient population studied, where shifts in phenytoin dosing and also changes in other factors (e.g., concomitant) were needed for patients with unstable clinical status.
The difference between PHTtotal/10
was also examined in relation to plasma albumin concentration (Figure , Additional file 1
: Figure S6A). The discrepancy between PHTfree
is most pronounced at low plasma albumin concentrations, where the ratio of PHTfree
would be expected to be higher. However, there are clearly many examples of marked discrepancies between PHTfree
even when the plasma albumin concentrations is within the age-specific reference range. PHTfree
is generally greater than PHTtotal/10
throughout all ages with examples of patients showing differences of > 2 mg/L evident throughout all age groups. The slope of the regression line in Figure was significantly different from 0 (i.e., null hypothesis of no relationship; 95% confidence interval; -2.37 to -1.90; P < 0.05). In contrast, the difference between PHTfree
shows little relationship with respect to plasma albumin concentration (Figure , Additional file 1
: Figure S6B), with the slope of the regression line in Figure showing no significant difference from 0 (95% confidence interval: -0.0649 to 0.0691).
Figure 3 Variation of total phenytoin, adjusted phenytoin, and free phenytoin with respect to plasma albumin concentration. (A) Variation of the difference between PHTtotal/10 and PHTfree with respect to plasma albumin concentration. All data is from chronologically (more ...)
The difference between PHTtotal/10
was also examined in relation to patient age (Figure , Additional file 1
: Figure S7A), which revealed no statistically significant difference of the slope of the regression line in Figure from 0 (95% confidence interval: -0.056 to 0.064). A similar finding was noted between the difference between PHTfree
(Figure , Additional file 1
: Figure S7B). The slope of the regression line in Figure showed no statistically significant difference from 0 (95% confidence interval: -0.053 to 0.057).
Figure 4 Variation of total phenytoin, adjusted phenytoin, and free phenytoin with respect to patient age. (A) Variation of the difference between PHTtotal/10 and PHTfree with respect to patient age. All data from chronologically first phenytoin measurements in (more ...)
A series of analyses were also done to try to understand what additional factors might influence how well PHTadj_free predicts PHTfree, using plots of the difference between PHTfree and PHTadj_free and various independent variables (Figure ; note that the statistical analyses in 5A-5D were four separate procedures). There is little influence of patient gender (Figure ) or days between albumin and PHTfree/PHTtotal measurements (Figure ) on the difference between PHTfree and PHTadj_free. There was, however, a significant effect of patient location at time of PHT measurements (Figure ), with the deviation between PHTfree and PHTadj_free highest in adult inpatient (non-ICU) and ICU units. In the inpatient settings, PHTfree was significantly greater than PHTadj_free. PHTfree - PHTadj_free did not vary significantly with the year in which the phenytoin measurements were performed (Figure ), suggesting that changes in clinical laboratory instrumentation and assays over the years of the retrospective analysis did not cause any changes in PHTfree, PHTtotal, or plasma albumin concentrations that might systematically impact the relation of PHTfree to PHTadj_free.
Figure 5 Variation of the difference of PHTfree and PHTadj_free with respect to various independent variables. (A) Distribution of the difference between PHTfree and PHTadj_free separated between males and females. Each dot represents a single timepoint of data (more ...)
Since its initial development, the Sheiner-Tozer equation has been widely used to assist therapeutic drug monitoring of phenytoin [14
]. It is even included in the MedMath module of Epocrates™ software (registered trademark of Epocrates, Inc., San Mateo, CA, USA,). Despite this widespread use, there has been considerable controversy over whether inaccuracies in the model justify its use or not. Some authors have argued that since PHTfree
levels may not be readily available, the adjustment in cases of known hypoalbuminemia provides better guidance in dosing than the total PHT level obtained in the usual assay used for therapeutic drug monitoring [9
]. In a population of "critically ill neurosurgical patients", Mlynarek et al. concluded that the Sheiner-Tozer equation provided "an unbiased, precise clinical estimate" in cases where the PHTfree
level "is unavailable or impractical" [17
]. For rural clinics in sub-Saharan Africa where malnutrition and AIDS are frequent, Fedler and Stewart concluded that the corrected value should be reported rather than the total phenytoin [9
]. On the other hand, two reports from university hospital settings concluded that because of the inaccuracies of the model, the Sheiner-Tozer equation should not be used [14
]. Other studies have recommended that PHTtotal
not be used at all, and that PHTfree
alone be used for drug monitoring of PHT [19
In our study, PHTadj_free
provides a better estimate of PHTfree
(relative to reference range) than PHTtotal
. Previous studies have focused on more limited sample sizes and patient populations [14
]. Our study included a population of mainly adults in the inpatient setting, including patients with refractory epilepsy and/or who were on multiple other anti-epileptic medications in addition to phenytoin. The linear regression lines relating PHTfree
have a slope close to 1 with only a slight negative bias (~0.2-0.3 mg/L) of PHTadj_free
relative to PHTfree
. However, we did demonstrate that the greatest bias between PHTfree
was seen in hospital inpatients, possibly due to other factors (e.g., concomitant drugs, organ failure) that can impact PHT pharmacokinetics.
In the university hospital setting, the cost to the clinical laboratory to perform the PHTfree
assay can be almost twice the cost of the PHTtotal
assay. In addition, the PHTfree
assay process includes an extra ultra-filtration step that requires centrifugation [13
]. Because of these extra steps, and for quality control, PHTfree
assays for a given day may be held and run in one or more batches during the day to limit labor-intensive steps. This contrasts with PHTtotal
levels that may run throughout the day on automated instrumentation without need for separate processing steps. Thus, there may be a delay in receiving the PHTfree
results compared to the availability of the PHTtotal
result on the same sample. In the case of a smaller hospital, where it may be longer until the batch of PHTfree
samples is accumulated and processed, or if PHTfree
levels are sent to a reference lab instead of done in house, the time disparity could be even greater. However, the added costs of the PHTfree
assay should be assessed in context with the risks of suboptimal phenytoin dosing (e.g., poor seizure control, toxicity, etc.).
Most of the "rule of thumb" equations in common use presume that the clinician is doing the calculation by hand or calculator, using a few readily available results to derive additional knowledge not directly reported. A modern electronic health record (EHR) with integrated DSS can provide added value for the clinician by doing these calculations automatically and posting the results, saving time and eliminating errors in calculation. One of the authors has taken this approach at the University of Pittsburgh Medical Center (UPMC) by developing and implementing rules to provide Anion Gap, Adjusted Sodium for hyperglycemia, and Adjusted Calcium for hypoalbuminemia, in addition to an Adjusted Phenytoin Rule. In all cases, interpretive data is attached to the result to further assist the clinician in decision making. Education of clinicians is important to emphasize the difference and limitations of both calculated and directly measured parameters.
There are several main limitations to the analysis presented in this paper. First, the patient population contains far more adults than pediatric patients. Consequently, the results are mostly applicable to adult patients. Future studies targeted at children, especially very young children, are needed. Second, the majority of datapoints arise from patients in inpatient units (including intensive care units) with lesser availability of datapoints arising from patients in outpatient clinics or the emergency room. However, the sample size of this study exceeds that of previous studies and contains a patient population that likely is similar to that analyzed by clinical laboratories at many academic medical centers, and has produced results comparable to other similar studies of academic medical center patient populations [14
]. Lastly, although nutritional status was not examined in detail, it is likely that most patients in the study were well-nourished and thus the findings are most applicable to other well-nourished populations.
Addressing the question of whether a better model than the Sheiner-Tozer equation could be implemented is the subject of additional analysis and development currently underway. The analysis reported here, along with the capabilities of the EHR and DSS, suggest additional possibilities to enhance the medical knowledge available to clinicians at the point of care. In addition to hypoalbuminemia, there are other factors that can influence binding of PHT and alter the free PHT fraction [14
]. With a "rule of thumb" type calculation, it is prohibitive to track these additional factors and perform the calculations they would entail by hand or using a non-programmable calculator. However, given the extensive data available electronically in the EHR and the logic capabilities of a modern DSS, it may become practical to implement far more complex models than those traditionally used in clinical practice. With knowledge discovery tools, the data available in the EHR database provide a substrate for increasingly sophisticated models. The development and testing of a model that predicts free phenytoin better than done by the Sheiner-Tozer model is the focus of additional research by the authors.