A recent publication by Ling J et al. has utilized eXtreme Gradient Boosting (XGBoost), a ML algorithm, to predict the 28-day mortality risk for sepsis patients based on 15 variables including RDW. The results of the paper showed that 28-day mortality in sepsis patients with thrombocytopenia was significantly higher than those without thrombocytopenia at the baseline (48.2% vs. 38.5%, respectively). SHAP interpretation of the XGBoost indicated that RDW was the second most important predictor of 28-day mortality in these patients following the SOFA scores. When comparing the subgroups of thrombocytopenia through AUC analysis, RDW was shown to be the most important predictor of 28-day mortality in thrombocytopenic patients with an area under the ROC of 0.646. An RDW of 16.05 displayed the best sensitivity and specificity for the prediction of mortality in these patients (70% and 57%, respectively) [27].
2.2. Drug-Induced Immune Thrombocytopenia
DITP is a common life-threatening complication seen in patients taking multiple drugs at the same time
[21][35]. This clinical syndrome is typically associated with severe bleeding that could ultimately result in death. There are many pathological mechanisms discussed in the literature to explain the causative mechanisms behind DITP
[21][22][23][24][35,36,37,38]. However, the most widely accepted mechanisms are direct bone marrow suppression by drugs and the development of drug-dependent antibodies (DDAbs) that activate platelets, ultimately leading to their depletion
[24][38]. The main dilemma with DITP is that it is challenging to diagnose clinically, and it is even more problematic to identify the causative drug
[25][39]. Moreover, the current experimental invitro methods of DITP diagnosis through DDAbs are unreliable, expensive, time-consuming, and are only available in a few platelet-specialized laboratories
[26][40]. Currently, the most effective approach for the treatment of patients with DITP is the cessation of the causative drug
[27][41].
2.2.1. Diagnosis of DITP
Linezolid is a synthetic antimicrobial used in the treatment of infections. Several studies have shown that linezolid could lead to linezolid associated thrombocytopenia (LAT) a type of DITP
[28][29][30][42,43,44]. The identification of LAT in patients taking linezolid treatment for infections may be difficult and time-consuming. To address this, Takahashi and colleagues conducted a study to create a classification tree model that predicts LAT in patients taking linezolid treatment. A total of 74 patients receiving linezolid treatment were retrospectively included in the study. LAT was defined as a 25% decline in platelet count from baseline. The baseline platelet count; age; total linezolid concentration; platelet count changes at 24, 48, 72, 96, and 120 h; creatinine clearance; and body weight were used as predictors for LAT in these patients. Binary decision trees were used to utilizing different combinations of the predictor variables to create tree models for LAT prediction. These trees are ML algorithms that classify observations by creating a sequence of binary questions. Binary questions are formed by creating the best possible splits for the data and repeats till further branching no longer improves the classification of observations. Trees were then pruned to avoid overfitting of the model to the learning data. The tree model with the lowest misclassification rate was taken as the final model
[31][13].
Another drug that is commonly associated with DITP is heparin. Approximately 1–3% of all patients treated with heparin develop HIT, and similarly to all other types of induced thrombocytopenia, it is a life-threatening condition
[32][45]. A major issue seen in the normal diagnostic approach for DITPs is that misdiagnosis is common due to lack of diagnostic data utilized
[33][34][46,47]. A study by Nilius and others utilized ML algorithms that integrate clinical and laboratory information to diagnose HIT more accurately than the traditional approach by the American Society of Hematology (ASH)
[35][15].
2.2.2. Predicting Drugs Causing DITP
Wang, B. et al. developed a several models to predict whether a drug could lead to DITP using seven different ML methods
[36][16]. A DITP dataset was collected from an online database, “platelets on the web”, which contained information on the compounds tested with DDAb
[37][48]. Compounds that had detectable DDAbs were classified as DITP toxicants (93) and those without DDAbs as non-toxicants (132). The dataset was then randomly divided in an 8:2 ratio into training and external validation sets, respectively. Support vector machine (SVM), k-nearest neighbor (k-NN), RF, naive bayes (NB), artificial neural network (ANN), adaptive boosting (AdaBoost), and XGBoost were used to produce binary classification models for DITP. Hyperparameters for each model was optimized by five-fold cross-validation, and the variance was reduced by 10× cross-validation. These models utilized six molecular fingerprints and three molecular descriptors to predict whether or not a drug can cause DITP
[36][16].
2.3. Hospital Acquired Thrombocytopenia
A common bleeding disorder following surgery is hospital acquired thrombocytopenia (HAT), This form of thrombocytopenia can be attributed to some of the previously discussed topics such as DIT or SAT. In a research paper conducted by Cheng and others, 7 ML models (GB, RF, logistic regression (LogR), XGBoost, multilayer perceptron, SVM, and k-NN) were created to predict patients at risk of HAT following surgery. Adult patients who were administered to the ICU following surgery were included in the study and divided in a 7:3 ratio into training and testing sets, respectively. Ten-fold validation was performed on the models generated. The
results
of
internal
validation
showed
that
the
RF
and
GB
models
outperformed
all
other
models
in
the
prediction
with
an
AUC
of
0.834
and
0.828,
respectively,
and
there
were
no
statistically
significant
differences
between
the
two
models.
Both
models
had
a
high
sensitivity
of
79.3%
and
73.6%,
respectively.
Specificity
for
the
models
were
79.1%
and
73.7%,
respectively [
46].
2.4. Immune Thrombocytopenia
Immune thrombocytopenia (ITP) is an autoimmune bleeding disorder
[38][49]. It is clinically characterized by low platelet count and the presence of autoantibodies to platelets
[39][50]. Diagnosis is typically made through the exclusion of all other causes of thrombocytopenia. ITP is considered a self-limiting disease in children where prognosis is good, and remission occurs easily. In adults however, ITP is chronic with a higher mortality rate
[40][41][51,52]. The pathogenesis of ITP is not well understood but involves the formation of IgG autoantibodies that target the glycoproteins IIb-IIIa on platelets. This results in the phagocytosis of these platelets hence leading to a drop in circulating platelet count and, ultimately, thrombocytopenia
[39][50].
2.4.1. Diagnosis
Kim, T. and others used a clinical database to create five ML models (RF, NB, LogR, SVM, and AdaBoost) to predict chronic immune thrombocytopenia in pediatric patients with ITP. A total of 969 pediatric patients with ITP were included in the study, of which, 332 had confirmed acute ITP and 253 with chronic ITP. Clinical (age, gender, race, ethnicity, presence of primary ITP) and laboratory variables (baseline platelet count, leukocyte count, lymphocyte count, eosinophil count, mean platelet volume, anti-nuclear antibody, immature platelet fraction, direct antiglobulin test, and immunoglobin levels) were used in the ML models to predict chronic ITP and 10-fold cross-validation was performed.
2.4.2. Prognosis
One of the main indicators for poor prognosis in patients with any form of thrombocytopenia is bleeding. A study by An, Z. Y. and others developed ML models to predict the risk of critical bleeds in patients with ITP. Data from eight centers in China were used to create this model utilizing nine predictor variables (Platelet count, age, onset of ITP, infection, type of ITP, bleeding of skin and mucosa, low platelet count, cardiovascular disease, and uncontrolled diabetes). Out of all the models, the RF model outperformed all others (AUC: 0.901 when externally validated)
[42][55].
3. Conclusions
ML algorithms can possibly provide a faster, reliable evaluation of patients with thrombocytopenia and, in some cases, may even outperform the current approaches for evaluating the diagnosis and prognosis of patients.