kyle.belanger/DHSC-Capstone
1
0
Fork 0
Code Issues 1 Pull requests Projects Releases Packages Wiki Activity Actions

updates

Browse source
This commit is contained in:
Kyle Belanger 2023-06-25 08:45:15 -04:00
parent 31c852f413
commit 5885dd08b3
8 changed files with 349 additions and 36 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
ML/2-modeling.R
View file

@ -140,6 +140,8 @@ gp2$ggsave(
)
ys$accuracy(class_test_results %>% tune::collect_predictions() ,truth = ft4_dia, estimate = .pred_class )
ys$sensitivity(class_test_results %>% tune::collect_predictions() ,truth = ft4_dia, estimate = .pred_class )
ys$specificity(class_test_results %>% tune::collect_predictions() ,truth = ft4_dia, estimate = .pred_class )
class_test_results %>%
workflows::extract_fit_parsnip() %>%

2
_quarto.yml
View file

@ -7,13 +7,13 @@ book:
date: "8/2/2022"
chapters:
- index.qmd
- abstract.qmd
- chapter1.qmd
- chapter2.qmd
- chapter3.qmd
- chapter4.qmd
- chapter5.qmd
- references.qmd
abstract: "This is a test to see what happens with this"

51
abstract.qmd Normal file
View file

@ -0,0 +1,51 @@
## Abstract {.unnumbered}
**Introduction**: This research study focuses on developing and testing
a machine learning algorithm to predict the FT4 result or diagnose hyper
or hypothyroidism in clinical chemistry. The goal is to bridge the gap
between hard-coded reflex testing and fully manual reflective testing
using machine learning algorithms. The significance of this study lies
in the increasing healthcare costs, where laboratory services contribute
significantly to medical decisions and budgets. By implementing
automated reflex testing with machine learning algorithms, unnecessary
laboratory tests can be reduced, resulting in cost savings and improved
efficiency in the healthcare system.
**Methods:** The study was performed using the Medical Information Mart
for Intensive Care (MIMIC) database for data collection. The database
consists of de-identified health-related data from critical care units.
Eighteen variables, including patient demographics and lab values, were
selected for the study. The data set was filtered based on specific
criteria, and an outcome variable was created to determine if the Free
T4 value was diagnostic. The data handling and modeling were performed
using R and R Studio. Regression and classification models were screened
using a random grid search to tune hyperparameters, and random forest
models were selected as the final models based on their performance. The
selected hyperparameters for both regression and classification models
are specified.
**Results:** The study analyzed a dataset of 11,340 observations,
randomly splitting it into a training set (9071 observations) and a
testing set (2269 observations) based on the Free T4 laboratory
diagnostic value stratification. Classification algorithms were used to
predict whether Free T4 would be diagnostic, achieving an accuracy of
0.796 and an AUC of 0.918. The model had a sensitivity of 0.632 and a
specificity of 0.892. The importance of individual analytes was
assessed, with TSH being the most influential variable. The study also
evaluated the predictability of Free T4 results using regression,
achieving a Root Mean Square Error (RMSE) of 0.334. The predicted
results had an accuracy of 0.790, similar to the classification model.
**Discussion:** The study found that the diagnostic value of Free T4 can
be accurately predicted 80% of the time using machine learning
algorithms. However, the model had limitations in terms of sensitivity,
with a false negative rate of 16% for elevated TSH results and 20% for
decreased TSH results. The model achieved a specificity of 89% but did
not meet the threshold for clinical deployment. The importance of
individual analytes was explored, revealing unexpected correlations
between TSH and hematology results, which could be valuable for future
algorithms. Real-world applications could use predictive models in
clinical decision-making systems to determine the need for Free T4 lab
tests based on predictions and patient signs and symptoms. However,
implementing such algorithms in existing laboratory information systems
poses challenges.

100
chapter3.qmd
View file

@ -2,11 +2,22 @@
## IRB
Based on the information you submitted for this project, the Campbell University Institutional Review Board (Campbell IRB) determined this submission is Not Human Subjects Research as defined by 45 CFR 46.102(e).
This study was submitted to the Cambell University Institutional Review
Board (Campbell IRB) . The study was determined to be Not Human Subjects
Research as defined by 45 CFR 46.102(e), and thus exempt from further
review by the IRB.
## Population and Data
This study used the Medical Information Mart for Intensive Care (MIMIC) database [@johnsonalistair]. MIMIC (Medical Information Mart for Intensive Care) is an extensive, freely-available database comprising de-identified health-related data from patients who were admitted to the critical care units of the Beth Israel Deaconess Medical Center. The database contains many different types of information, but only data from the patients and laboratory events table are used in this study. The study uses version IV of the database, comprising data from 2008 - 2019.
This study used the Medical Information Mart for Intensive Care (MIMIC)
database [@johnsonalistair]. MIMIC (Medical Information Mart for
Intensive Care) is an extensive, freely-available database comprising
de-identified health-related data from patients who were admitted to the
critical care units of the Beth Israel Deaconess Medical Center. The
database contains many different types of information, but only data
from the patients and laboratory events table are used in this study.
The study uses version IV of the database, comprising data from 2008 -
2019.
## Data Variables and Outcomes
@ -19,19 +30,37 @@ source(here::here("ML","1-data-exploration.R"))
```
A total of 18 variables were chosen for this study. The age and gender of the patient were pulled from the patient table in the MIMIC database. While this database contains some additional demographic information, it is incomplete and thus unusable for this study. 15 lab values were selected for this study, this includes:
A total of 18 variables were chosen for this study. The age and gender
of the patient were pulled from the patient table in the MIMIC database.
While this database contains some additional demographic information, it
is incomplete and thus unusable for this study. 15 lab values were
selected for this study, this includes:
- **BMP**: BUN, bicarbonate, calcium, chloride, creatinine, glucose, potassium, sodium
- **BMP**: BUN, bicarbonate, calcium, chloride, creatinine, glucose,
potassium, sodium
- **CBC**: Hematocrit, hemoglobin, platelet count, red blood cell count, white blood cell count
- **CBC**: Hematocrit, hemoglobin, platelet count, red blood cell
count, white blood cell count
- TSH
- Free T4
The unique patient id and chart time were also retained for identifying each sample. Each sample contains one set of 15 lab values for each patient. Patients may have several samples in the data set run at different times. Rows were retained as long as they had less than three missing results. These missing results can be filled in by imputation later in the process. Samples were also filtered for those with TSH above or below the reference range of 0.27 - 4.2 uIU/mL. These represent samples that would have reflexed for Free T4 testing. After filtering, the final data set contained `r nrow(ds1)` rows.
The unique patient id and chart time were also retained for identifying
each sample. Each sample contains one set of 15 lab values for each
patient. Patients may have several samples in the data set run at
different times. Rows were retained as long as they had less than three
missing results. These missing results can be filled in by imputation
later in the process. Samples were also filtered for those with TSH
above or below the reference range of 0.27 - 4.2 uIU/mL. These represent
samples that would have reflexed for Free T4 testing. After filtering,
the final data set contained `r nrow(ds1)` rows.
Once the final data set was collected, an additional column was created for the outcome variable to determine if the Free T4 value was diagnostic. This outcome variable was used for building classification models. The classification variable was not used in regression models. @tbl-outcome_var shows how the outcomes were added
Once the final data set was collected, an additional column was created
for the outcome variable to determine if the Free T4 value was
diagnostic. This outcome variable was used for building classification
models. The classification variable was not used in regression models.
@tbl-outcome_var shows how the outcomes were added
| TSH Value | Free T4 Value | Outcome |
|---------------|---------------|---------------------|
@ -42,7 +71,14 @@ Once the final data set was collected, an additional column was created for the
: Outcome Variable {#tbl-outcome_var}
. @tbl-data_summary shows the summary statistics of each variable selected for the study. Each numeric variable is listed with the percent missing, median, and interquartile range (IQR). The data set is weighted toward elevated TSH levels, with 80% of values falling into that category. Glucose and Calcium have several missing values at `r gtsummary::inline_text(summary_tbl, variable = GLU, column = n)` and `r gtsummary::inline_text(summary_tbl, variable = CA, column = n)`, respectively.
. @tbl-data_summary shows the summary statistics of each variable
selected for the study. Each numeric variable is listed with the percent
missing, median, and interquartile range (IQR). The data set is weighted
toward elevated TSH levels, with 80% of values falling into that
category. Glucose and Calcium have several missing values at
`r gtsummary::inline_text(summary_tbl, variable = GLU, column = n)` and
`r gtsummary::inline_text(summary_tbl, variable = CA, column = n)`,
respectively.
```{r}
#| label: tbl-data_summary
@ -54,19 +90,41 @@ summary_tbl %>% gtsummary$as_kable()
## Data Inspection
By examining @tbl-data_summary several important data set characteristics quickly come to light without explanation. The median age across the data set, as a whole, is quite similar, with a median age across all categories of 62.5. Females are better represented in the data set, with higher percentages in all categories. Across all categories, the median values for each lab result are pretty similar. The expectation for this is Red Blood cells, which show more considerable variation across the various categories.
By examining @tbl-data_summary several important data set
characteristics quickly come to light without explanation. The median
age across the data set, as a whole, is quite similar, with a median age
across all categories of 62.5. Females are better represented in the
data set, with higher percentages in all categories. Across all
categories, the median values for each lab result are pretty similar.
The expectation for this is Red Blood cells, which show more
considerable variation across the various categories.
![Distribution of Variables](figures/distrubution_histo){#fig-distro_histo}
![Distribution of
Variables](figures/distrubution_histo){#fig-distro_histo}
When examining @fig-distro_histo, many clinical chemistry values do not show a standard distribution. However, the hematology results typically do appear to follow a standard distribution. While not a problem for most tree-based classification models, many regression models perform better with standard variables. Standardizing variables provides a common comparable unit of measure across all the variables [@boehmke2020]. Since lab values do not contain negative numbers, all numeric values will be log-transformed to bring them to normal distributions.
When examining @fig-distro_histo, many clinical chemistry values do not
show a standard distribution. However, the hematology results typically
do appear to follow a standard distribution. While not a problem for
most tree-based classification models, many regression models perform
better with standard variables. Standardizing variables provides a
common comparable unit of measure across all the variables
[@boehmke2020]. Since lab values do not contain negative numbers, all
numeric values will be log-transformed to bring them to normal
distributions.
![Variable Correlation Plot](figures/corr_plot){#fig-corr_plot}
@fig-corr_plot shows a high correlation between Hemoglobin, hematocrit, and Red Blood Cell values (as expected). While high correlation does not lead to model issues, it can cause unnecessary computations with little value. However, due to the small number of variables, the computation burden is not expected to cause delays, and thus the variables will not be removed.
@fig-corr_plot shows a high correlation between Hemoglobin, hematocrit,
and Red Blood Cell values (as expected). While high correlation does not
lead to model issues, it can cause unnecessary computations with little
value. However, due to the small number of variables, the computation
burden is not expected to cause delays, and thus the variables will not
be removed.
## Data Tools
All data handling and modeling were performed using R and R Studio. The current report was rendered in the following environment.
All data handling and modeling were performed using R and R Studio. The
current report was rendered in the following environment.
```{r}
#| label: tbl-platform-info
@ -120,7 +178,14 @@ knitr::kable(
## Model Selection
Both classification and regression models were screened using a random grid search to tune hyperparameters. The models were tested against the training data set to find the best-fit model. @fig-reg-screen shows the results of the model screening for regression models, using root mean square error (RMSE) as the ranking method. Random Forest models and boosted trees performed similarly and were selected for further testing. A full grid search was performed on both models, with a Random Forest model as the final selection. The final hyperparameters selected were:
Both classification and regression models were screened using a random
grid search to tune hyperparameters. The models were tested against the
training data set to find the best-fit model. @fig-reg-screen shows the
results of the model screening for regression models, using root mean
square error (RMSE) as the ranking method. Random Forest models and
boosted trees performed similarly and were selected for further testing.
A full grid search was performed on both models, with a Random Forest
model as the final selection. The final hyperparameters selected were:
- mtry: 8
@ -130,7 +195,12 @@ Both classification and regression models were screened using a random grid sear
![Regression Model Screen](figures/reg_screen){#fig-reg-screen}
@fig-class-screen shows the results of the model screen for classification models using accuracy as the ranking method. As with regression models, boosted trees and random forest models performed the best. After completing a full grid search of both model types, a random forest model was again chosen as the final model. The final hyperparameters for the model selected were:
@fig-class-screen shows the results of the model screen for
classification models using accuracy as the ranking method. As with
regression models, boosted trees and random forest models performed the
best. After completing a full grid search of both model types, a random
forest model was again chosen as the final model. The final
hyperparameters for the model selected were:
- mtry: 8

62
chapter4.qmd
View file

@ -8,7 +8,13 @@ load(here::here("figures", "strata_table.Rda"))
```
The final data set used for this analysis consisted of 11,340 observations. All observations contained a TSH and Free T4 result and less than three missing results from all other analytes selected for the study. The dataset was then randomly split into a training set containing 9071 observations and a testing set containing 2269 observations. The data was split using stratification of the Free T4 laboratory diagnostic value. @tbl-strata shows the split percentages.
The final data set used for this analysis consisted of 11,340
observations. All observations contained a TSH and Free T4 result and
less than three missing results from all other analytes selected for the
study. The dataset was then randomly split into a training set
containing 9071 observations and a testing set containing 2269
observations. The data was split using stratification of the Free T4
laboratory diagnostic value. @tbl-strata shows the split percentages.
```{r}
#| label: tbl-strata
@ -19,26 +25,66 @@ strata_table %>% knitr::kable()
```
First, the report shows the ability of classification algorithms to predict whether Free T4 will be diagnostic, with the prediction quality measured by Area Under Curve (AUC) and accuracy. Data regarding the importance association between each predictor analyte and the Free T4 Diagnostic value is then presented. Finally, data is presented with the extent to which FT4 can be predicted by examining the correlation statistics denoting the relationship between measured and predicted Free T4 values.
First, the report shows the ability of classification algorithms to
predict whether Free T4 will be diagnostic, with the prediction quality
measured by Area Under Curve (AUC) and accuracy. Data regarding the
importance association between each predictor analyte and the Free T4
Diagnostic value is then presented. Finally, data is presented with the
extent to which FT4 can be predicted by examining the correlation
statistics denoting the relationship between measured and predicted Free
T4 values.
## Predictability of Free T4 Classifications
In clinical decision-making, a key consideration in interpreting numerical laboratory results is often just whether the results fall within the normal reference range [@luo2016]. In the case of Free T4 reflex testing, the results will either fall within the normal range indicating the Free T4 is not diagnostic of Hyper or Hypo Throydism, or they will fall outside those ranges indicating they are diagnostic. The final model achieved an accuracy of 0.796 and an AUC of 0.918. @fig-roc_curve provides ROC curves for each of the four outcome classes.
In clinical decision-making, a key consideration in interpreting
numerical laboratory results is often just whether the results fall
within the normal reference range [@luo2016]. In the case of Free T4
reflex testing, the results will either fall within the normal range
indicating the Free T4 is not diagnostic of Hyper or Hypo Throydism, or
they will fall outside those ranges indicating they are diagnostic. The
final model achieved an accuracy of 0.796 and an AUC of 0.918.
@fig-roc_curve provides ROC curves for each of the four outcome classes.
The same model achieved a sensitivity of 0.632 and specificity of 0.892
![ROC curves for each of the four outcome classes](figures/roc_curve_class){#fig-roc_curve}
![ROC curves for each of the four outcome
classes](figures/roc_curve_class){#fig-roc_curve}
@fig-conf-matrix-class shows the confusion matrix of the final testing data. Of the 2269 total results, 1805 were predicted correctly, leaving 464 incorrectly predicted results. Of the incorrectly predicted results, 72 results predicted a diagnostic Free T4 when the correct result was non-diagnostic. 392 of the incorrectly predicted results were predicted as non-diagnostic when the correct result was diagnostic.
@fig-conf-matrix-class shows the confusion matrix of the final testing
data. Of the 2269 total results, 1805 were predicted correctly, leaving
464 incorrectly predicted results. Of the incorrectly predicted results,
72 results predicted a diagnostic Free T4 when the correct result was
non-diagnostic. 392 of the incorrectly predicted results were predicted
as non-diagnostic when the correct result was diagnostic.
![Final Model Confusion Matrix](figures/conf_matrix_class){#fig-conf-matrix-class}
![Final Model Confusion
Matrix](figures/conf_matrix_class){#fig-conf-matrix-class}
## Contributions of Individual Analytes
Understanding how an ML model makes predictions helps build trust in the model and is the fundamental idea of the emerging field of interpretable machine learning (IML) [@greenwell2020]. @fig-vip-class shows the importance of features in the final model. Importance can be defined as the extent to which a feature has a "meaningful" impact on the predicted outcome [@laan2006]. As expected, TSH is the leading variable in importance rankings, leading all other variables by over 2000's points. The following three variables are all parts of a Complete Blood Count (CBC), followed by the patient's glucose value.
Understanding how an ML model makes predictions helps build trust in the
model and is the fundamental idea of the emerging field of interpretable
machine learning (IML) [@greenwell2020]. @fig-vip-class shows the
importance of features in the final model. Importance can be defined as
the extent to which a feature has a "meaningful" impact on the predicted
outcome [@laan2006]. As expected, TSH is the leading variable in
importance rankings, leading all other variables by over 2000's points.
The following three variables are all parts of a Complete Blood Count
(CBC), followed by the patient's glucose value.
![Variable Importance Plot](figures/vip_class){#fig-vip-class}
## Predictability of Free T4 Results (Regression)
Today, it has become widely accepted that a more sound approach to assessing model performance is to assess the predictive accuracy via loss functions. Loss functions are metrics that compare the predicted values to the actual value (the output of a loss function is often referred to as the error or pseudo residual) [@boehmke2020]. The loss function used to evaluate the final model was selected as the Root Mean Square Error, and the final testing data achieved an RMSE of 0.334. @fig-reg-pred shows the plotted results. The predicted results were also used to add the diagnostic classification of Free T4. These results achieved an accuracy of 0.790, and thus very similar to the classification model.
Today, it has become widely accepted that a more sound approach to
assessing model performance is to assess the predictive accuracy via
loss functions. Loss functions are metrics that compare the predicted
values to the actual value (the output of a loss function is often
referred to as the error or pseudo residual) [@boehmke2020]. The loss
function used to evaluate the final model was selected as the Root Mean
Square Error, and the final testing data achieved an RMSE of 0.334.
@fig-reg-pred shows the plotted results. The predicted results were also
used to add the diagnostic classification of Free T4. These results
achieved an accuracy of 0.790, and thus very similar to the
classification model.
![Regression Predictions Plot](figures/reggression_pred){#fig-reg-pred}

147
chapter5.qmd
View file

@ -1,31 +1,154 @@
# Discussion
Intro Paragraph - In progress<!--# Write after I write everything else -->
## Summary of Results
The findings of this study indicate that within another commonly ordered laboratory testing, the diagnostic value of Free T4 can be predicted accurately 80% of the time. While examining only the elevated TSH results, the algorithm had a false positive rate of 2% and a false negative rate of 16%. In the original data, 76% of the time, the result was non-diagnostic for Hypo-Thryodism. For the decreased TSH results, the algorithm had a false positive rate of 8% and a false negative rate of 20%. In the original data, 67% of the time, the result was non-diagnostic for Hyper-Thryodism.
The findings of this study indicate that within another commonly ordered
laboratory testing, the diagnostic value of Free T4 can be predicted
accurately 80% of the time. While examining only the elevated TSH
results, the algorithm had a false positive rate of 2% and a false
negative rate of 16%. In the original data, 76% of the time, the result
was non-diagnostic for Hypo-Thryodism. For the decreased TSH results,
the algorithm had a false positive rate of 8% and a false negative rate
of 20%. In the original data, 67% of the time, the result was
non-diagnostic for Hyper-Thryodism.
While TSH was expected to be the most important variable in building random forest models, it was entirely unexpected that the following three values would be Hematology results. In the clinical laboratory, TSH and CBCs are often run on different analyzers and in other departments. Finding this slight correlation could be valuable to building further algorithms.
While the model achieved an overall accuracy of 80%, it struggled to
identify positives with a sensitivity of only 63%. However, the model
did achieve a specificity of 89%. Sensitivity refers to a test's ability
to designate an individual with the disease as positive. A highly
sensitive test means few false negative results, and thus fewer disease
cases are missed. The specificity of a test is its ability to designate
an individual who does not have a disease as negative. A highly specific
test means that there are few false positive results. It may not be
feasible to use a test with low specificity for screening since many
people without the disease will screen positive and potentially receive
unnecessary diagnostic procedures [@newyorkstatedepartmentofhealth].
In a study by Xu et al., a machine learning model was used to predict
laboratory test results as normal or abnormal to identify low-yield,
repetitive laboratory tests [-@xu2019]. Their group performed a
multi-site study of nearly 200,000 inpatient laboratory testing orders
to identify the most repetitive laboratory tests and then attempted to
predict each one. They achieved an AUROC of \> 90% for 20 common
laboratory tests, including sodium, hemoglobin, and lactate
dehydrogenase. They proposed a sensitive decision threshold of a
negative predictive value of 95% to power a clinical decision support
tool aimed at reducing low-yield, repetitive testing [@xu2019]. No other
published studies exist in the clinical laboratory with a proposed value
for the success of a machine learning model. If using the 95%
specificity threshold, the current model does not achieve the result
necessary to be considered final.
While TSH was expected to be the most important variable in building
random forest models, it was entirely unexpected that the following
three values would be Hematology results. In the clinical laboratory,
TSH and CBCs are often run on different analyzers and in other
departments. Finding this slight correlation could be valuable to
building further algorithms. The currently available literature states
TSH and fT4 have a complex, nonlinear relationship, such that small
changes in fT4 result in relatively large changes in TSH [@plebani2020].
However, no currently available literature explores a relationship
between TSH and any of the CBC tests. These small changes between FT4
and TSH may be explained if this link can be expanded. While this study
only focuses on high-level CBC testing, most automated CBC analyzers can
run many more tests, which could be used in the development of future
algorithms.
## Real World Applications
While the current algorithm did not quite achieve an accuracy ready for deployment, it is hypothesized that a system like this could be implemented in clinical decision-making systems. As stated previously, current practice is a physician (or other care providers) orders a TSH, and if the value is outside laboratory-established reference ranges, the Free T4 is added on. In the current study database, this reflex testing was non-diagnostic 76% of the time for elevated TSH values and 67% for decreased TSH values. Using clinical decision support first to predict whether the Free T4 would be diagnostic, the care provider can use this prediction and other patient signs and symptoms to determine if running a Free T4 lab test is needed.
While the current algorithm did not quite achieve an accuracy ready for
deployment, it is hypothesized that a system like this could be
implemented in clinical decision-making systems. As stated previously,
current practice is a physician (or other care providers) orders a TSH,
and if the value is outside laboratory-established reference ranges, the
Free T4 is added on. In the current study database, this reflex testing
was non-diagnostic 76% of the time for elevated TSH values and 67% for
decreased TSH values. Using clinical decision support first to predict
whether the Free T4 would be diagnostic, the care provider can use this
prediction and other patient signs and symptoms to determine if running
a Free T4 lab test is needed.
Similarly to Luo et al., the idea that the diagnostic information offered by Free T4 often duplicates what other diagnostic tests provide suggests a notion of "informationally" redundant testing [-@luo2016]. It is speculated that informationally redundant testing occurs in various diagnostic settings and diagnostic workups. It is much more frequent than the more traditionally defined and narrowly framed notion of redundant testing, which most often includes unintended duplications of the same or similar tests. Under this narrow definition, redundant laboratory testing is estimated to waste more than \$5 billion annually in the United States, potentially dwarfed by the waste from informationally redundant testing [@luo2016]. However, since Free T4 and all other tests used in this study are performed on automated instruments, the cost savings to the lab and patient may be minimal.
Similarly to Luo et al., the idea that the diagnostic information
offered by Free T4 often duplicates what other diagnostic tests provide
suggests a notion of "informationally" redundant testing [-@luo2016]. It
is speculated that informationally redundant testing occurs in various
diagnostic settings and diagnostic workups. It is much more frequent
than the more traditionally defined and narrowly framed notion of
redundant testing, which most often includes unintended duplications of
the same or similar tests. Under this narrow definition, redundant
laboratory testing is estimated to waste more than \$5 billion annually
in the United States, potentially dwarfed by the waste from
informationally redundant testing [@luo2016]. However, since Free T4 and
all other tests used in this study are performed on automated
instruments, the cost savings to the lab and patient may be minimal.
As Rabbani et al. study showed, Machine Learning in the Clinical Laboratory is an emerging field. However, few existing studies relate to predicting laboratory values based on other results [-@rabbani2022]. The few studies that do exist follow a similar premise. All are trying to reduce redundant laboratory testing, thus lowering the patient's cost.
As Rabbani et al. study showed, Machine Learning in the Clinical
Laboratory is an emerging field. However, few existing studies relate to
predicting laboratory values based on other results [-@rabbani2022]. The
few studies that do exist follow a similar premise. All are trying to
reduce redundant laboratory testing, thus lowering the patient's cost.
## Study Limitations
While the MIMIC-IV database allowed for a first run of the study, it does suffer from some issues compared to other patient results. The MIMIC-IV database only contains results from ICU patients. Thus the result may not represent normal results for patients typically screened for hyper or hypothyroidism. In a study by Tyler et al., they found that laboratory value ranges from critically ill patients deviate significantly from those of healthy controls [-@tyler2018]. In their study, distribution curves based on ICU data, have differed considerably from the standard hospital range (mean \[SD\] overlapping coefficient, 0.51 \[0.32-0.69\]) [@tyler2018]. The data ranges from 2008 to 2019. During this time, there could have been several unknown laboratory changes. Often laboratories change methods, reference ranges, or even vendors. None of this data is available in the MIMIC database. A change in method or vendor could cause a shift in results, thus causing the algorithm to assign incorrect outcomes.
While the MIMIC-IV database allowed for a first run of the study, it
does suffer from some issues compared to other patient results. The
MIMIC-IV database only contains results from ICU patients. Thus the
result may not represent normal results for patients typically screened
for hyper or hypothyroidism. In a study by Tyler et al., they found that
laboratory value ranges from critically ill patients deviate
significantly from those of healthy controls [-@tyler2018]. In their
study, distribution curves based on ICU data, have differed considerably
from the standard hospital range (mean \[SD\] overlapping coefficient,
0.51 \[0.32-0.69\]) [@tyler2018]. The data ranges from 2008 to 2019.
During this time, there could have been several unknown laboratory
changes. Often laboratories change methods, reference ranges, or even
vendors. None of this data is available in the MIMIC database. A change
in method or vendor could cause a shift in results, thus causing the
algorithm to assign incorrect outcomes.
The dataset also sufferers from incompleteness. Due to the fact the database was not explicitly designed for this study, many patients do not have complete sets of lab results. The study also had to pick and choose lab tests to allow for as many groups of TSH and Free T4 results as possible. For instance, in a study by Luo et al., a total of 42 different lab tests were selected for a Machine Learning study, compared to only 16 selected for this study [-@luo2016]. The patient demographic data also suffered from the same incompleteness. Due to this fact, only the age and gender of the patient were used in developing the algorithm. An early study by Schectman et al. found the mean TSH level of Blacks was 0.4 (SE .053) mU/L lower than that for Whites after age and sex adjustment, race explaining 6.5 percent of the variation in TSH levels [-@schectman1991]. This variation in results should potentially be included in developing a future algorithm. However, as it stands, the current data set has incomplete data for patient race and ethnicity.
The dataset also sufferers from incompleteness. Due to the fact the
database was not explicitly designed for this study, many patients do
not have complete sets of lab results. The study also had to pick and
choose lab tests to allow for as many groups of TSH and Free T4 results
as possible. For instance, in a study by Luo et al., a total of 42
different lab tests were selected for a Machine Learning study, compared
to only 16 selected for this study [-@luo2016]. The patient demographic
data also suffered from the same incompleteness. Due to this fact, only
the age and gender of the patient were used in developing the algorithm.
An early study by Schectman et al. found the mean TSH level of Blacks
was 0.4 (SE .053) mU/L lower than that for Whites after age and sex
adjustment, race explaining 6.5 percent of the variation in TSH levels
[-@schectman1991]. This variation in results should potentially be
included in developing a future algorithm. However, as it stands, the
current data set has incomplete data for patient race and ethnicity.
As Machine learning algorithms become more and more powerful, it is additionally vital from an infrastructure standpoint to have the processing power capable of handling the algorithms. This becomes even more important in an attempt to put the algorithm into practice, as the computer must be able to process results in mere milliseconds.
As Machine learning algorithms become more and more powerful, it is
additionally vital from an infrastructure standpoint to have the
processing power capable of handling the algorithms. This becomes even
more important in an attempt to put the algorithm into practice, as the
computer must be able to process results in mere milliseconds.
## Future Studies
While the current algorithm is not quite ready for production use, it does lead to many promising ideas. The first step to further develop this algorithm would be collecting data on non-ICU patients. The idea would be gathering data on patients much closer to those screened for Hypo and Hyper-Thyrodism. With data closer to normal, the optimal hyperparameters could continue to be tweaked, as well as training the model with this data. There could also be a reason to try and test the current algorithm with different patient data to assess performance. This would be similar to what Li et al. performed with their study to identify unnecessary laboratory tests [-@li2022]. After developing their algorithm on the MIMIC-III database, they gathered data from Memorial Hermann Hospital in Houston, Texas. However, their algorithm was designed for ICU patients in this study, so this was a more direct performance comparison. In the case of this study, the algorithm was intended more as a proof of concept than are production-ready idea.
While the current algorithm is not quite ready for production use, it
does lead to many promising ideas. The first step to further develop
this algorithm would be collecting data on non-ICU patients. The idea
would be gathering data on patients much closer to those screened for
Hypo and Hyper-Thyrodism. With data closer to normal, the optimal
hyperparameters could continue to be tweaked, as well as training the
model with this data. There could also be a reason to try and test the
current algorithm with different patient data to assess performance.
This would be similar to what Li et al. performed with their study to
identify unnecessary laboratory tests [-@li2022]. After developing their
algorithm on the MIMIC-III database, they gathered data from Memorial
Hermann Hospital in Houston, Texas. However, their algorithm was
designed for ICU patients in this study, so this was a more direct
performance comparison. In the case of this study, the algorithm was
intended more as a proof of concept than are production-ready idea.
One of the most challenging parts of this study and any machine learning in the clinical laboratory is implementing it after the fact. Developing an algorithm that can predict laboratory testing is just half the idea. Many current laboratory information systems would be unable to handle this type of clinical decision-making system, as this would be much outside the expected behavior of these systems.
One of the most challenging parts of this study and any machine learning
in the clinical laboratory is implementing it after the fact. Developing
an algorithm that can predict laboratory testing is just half the idea.
Many current laboratory information systems would be unable to handle
this type of clinical decision-making system, as this would be much
outside the expected behavior of these systems.

BIN
extras/irbletter.docx Normal file
View file

Binary file not shown.

21
references.bib
View file

@ -410,3 +410,24 @@ PMCID: PMC6324400}
note = {PMID: 2003636
PMCID: PMC1405055}
}
@misc{newyorkstatedepartmentofhealth,
title = {Disease Screening - Statistics Teaching Tools - New York State Department of Health},
author = {New York State Department of Health, },
url = {https://www.health.ny.gov/diseases/chronic/discreen.htm}
}
@article{xu2019,
title = {Prevalence and Predictability of Low-Yield Inpatient Laboratory Diagnostic Tests},
author = {Xu, Song and Hom, Jason and Balasubramanian, Santhosh and Schroeder, Lee F. and Najafi, Nader and Roy, Shivaal and Chen, Jonathan H.},
year = {2019},
month = {09},
date = {2019-09-11},
journal = {JAMA Network Open},
pages = {e1910967},
volume = {2},
number = {9},
doi = {10.1001/jamanetworkopen.2019.10967},
url = {https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2749559},
langid = {en}
}