In this section, we explain the data set, data preprocessing methods, blood transfusion prediction, and the proposed nano fuzzy alarming system for blood transfusion requirements.DatasetWe use a dataset of 98 samples of blood cancer patients including 61 features of demographic, clinical, and laboratory data from Firozgar Hospital. Since the research did not impact clinical care and all information was deidentified, the requirement for individual patient consent was waived by review boards of the Iran University of Medical Sciences (IUMS). All procedures were performed in accordance with relevant guidelines. All experimental protocols were approved by the review boards of IUMS. The data has been de-identified and is not subject to HIPAA Privacy Rule restrictions. The used features which were selected after performing multivariate analysis and approval of an expert are presented in Table 1 in three categories including demographic information, clinical features, and three sets of laboratory data.
Table 1 Selected features for predicting blood transfusions.Data preprocessingHere we use vitals and laboratory values for each patient. We exclude values with more than 90% missingness including BMI, chemotherapy intervals, chloride, anion gap, total iron, iron-binding capacity, lactate, and transferrin. For those values with more than 50% missingness, we consider 0 as an indicator for missingness and 1 as those which are present. Then we use the median imputation with those values that have a missingness indicator. To have a similar protocol across patients in the values of heart rate, systolic blood pressure, and diastolic blood pressure, we fix the first timepoint of each recorded data at the first recording. We then normalize the continuous data before training to avoid the effect of much larger values to the rest of the variables.We consider 1 as representative of a man and 2 as a woman in gender features. For the history of disease including diabetes, heart, respiratory, and Chronic kidney in addition to sepsis, active infection, presence of active bleeding, and fever, we use 1 as an indicator of having the disease and 0 as not having the disease. We consider 1–4 for four types of blood cancer including Acute Lymphoblastic Leukemia (ALL), Acute Myeloid Leukemia (AML), Chronic Lymphocytic Leukemia (CLL), and Chronic Myelogenous Leukemia (CML). Moreover, for the smoking and alcohol consumption in addition to chemotherapy, antibiotic injection, and blood transfusion, we use 1 as an indicator of applying and 0 as not applying. Finally, we use 0 for negative and 1 for positive Troponin T values. In the training set 7%, and in the test set 4% are labeled as receiving a packed red blood cell.To decrease the effect of correlated variables in the neural network variables, we use a Pearson’s correlation matrix as shown in Fig. 1 for geographical data, Fig. 2 for clinical data, Figs. 3, 4, and 5 for laboratory data to validate that input variables are relatively uncorrelated. As the result shows they are uncorrelated. In Fig. 1, all demographic data except age are Boolean variables, and the rest of the variables are continuous real values.Figure 1Pearson’s correlation matrix for geographical data.Figure 2Pearson’s correlation matrix for clinical data.Figure 3Pearson’s correlation matrix for the first set of laboratory data.Figure 4Pearson’s correlation matrix for the second set of laboratory data.Figure 5Pearson’s correlation matrix for third sets of laboratory data.We then detect the relevant variables. In Fig. 1 sepsis, active infection, and fever are related. DBP and SBP are related in Fig. 2. In Fig. 3 Hgb and Hct are related to platelet. Also, MCH is related to MCHC, and TCD is related to BE. In Fig. 4, BUN and creatinine are relevant. Also, PTT and PT are relevant. In Fig. 5, CK and CK-MB are relevant to Troponin T, CK is related to CK-MB, direct BR is related to TBIL, and finally, ALP and AST are related to ALT.Distinguishable parameters selection in blood transfusionWe first analyze data to find the distinguishable parameters in blood transfusion requirements. Hence, we illustrate the scatterplot of data depending on two categories of blood transfusion and not blood transfusion in Figs. 6, 7, 8, and 9. Accordingly, SBP in Fig. 6 and platelet, Hct, Hgb, and the variables of SpO2, PaO2, pH, TCD, and BE in the presence of Hct and Hgb in Fig. 7 are selected as the most distinguishable parameters. Furthermore, in Fig. 8, creatinine, BUN, and PTT are selected, and finally, in Fig. 9 there are no distinguishable variables.Figure 6Scatterplot of clinical data for the two classes of blood transfusion.Figure 7Scatterplot of laboratory data for the two classes of blood transfusion.Figure 8Scatterplot of second sets of laboratory data for the two classes of blood transfusion.Figure 9Scatterplot of third sets of laboratory data for the two classes of blood transfusion.Considering the analyzed data, we then use these distinguishable variables as a possible artificial intelligence (AI)-based biomarker for blood transfusion requirements. Hence, in this paper, we sense the changes in the values of these biomarkers using nanomachine to determine the right time for blood transfusion. For this, we first perform the prediction and then utilize the data for proposing a nano-alarming system for blood transfusion requirements.Blood transfusion predictionTo perform the prediction, we use a type of Recurrent Neural Network (RNN) as Long Short Term Memory (LSTM). LSTM can handle information that is passed between subsequent time iterations. In this modeling, \(x\left(0\right)\), \(x\left(1\right)\), …, \(x\left(T-1\right)\) where represents the input variables at the beginning of each 24 h and \(\widehat{y}\left(0\right)\), \(\widehat{y}\left(1\right)\),…,\(\widehat{y}\left(T\right)\), where \(\widehat{y}\left(T\right)\in \left[0\text{,}1\right]\) is the output that predicts the blood transfusion. We consider LogSoftmax for the last layer to get the log-probabilities of the output including \({\text{p}}\left({1}\right)\text{, p}\left({2}\right)\text{, . . . , p}\left({\text{T}}\right)\) in which \(p\left(t\right)\) is the log-probability of \(\widehat{y}\) for the two classes.Nano-alarming system for blood transfusion requirementIn this section, we propose a feasibility study on nano fuzzy alarming system (NFABT) for blood transfusion requirement. Previously, inspired by red blood cells24 and cancer cell scaffold25, we proposed bio-inspired nanomachines for cancer drug delivery26, and oxygen deficiency cases7. Here, utilizing red blood cell-inspired nanomachines, we propose a nano-alarming system for blood transfusion. For this, each nanomachine is equipped with fuzzy basis functions (FBFs), and the overall fuzzy system is calculated using a swarm of FBFs. Figure 10 shows the proposed NFABT method.Figure 10Proposed nano fuzzy alarming system (NFABT) for blood transfusion requirement.Considering the analysis in “Data preprocessing” we select three parameters of Hgb, PaO2, and pH value for possible inputs of FBFs in NFABT. These values are sensed using embedded bio nanosensors for blood in each nanomachine. The output of this system is the state of blood regarding the need for transfusion and the overall system alarms the transfusion requirement.Accordingly, the inputs of Hgb, PaO2, and pH and the output of the blood state as the marker for blood transfusion requirement in the universe of discourse of [7,16.6], [40,80], [6.3,7.8], and [0,1], respectively. Considering experts’ suggestions and clinical guidelines, we define four MFs including {Very Low, Low, Normal, high}, {Normal, Mild Hypoxemia, Moderate Hypoxemia, Severe hypoxemia}, and {Acidosis, Normal, Alkalosis} for inputs and two MFs including {Low, Normal} for output. The centers of the MFs of the inputs and outputs are {8,10.25,14.1,16.6}, {30,47.5,65.75,90}, {7.35,7.4,7.45}, and {0.25,0.75}, respectively. We also have \({4}^{2}\times {2}^{1}=32\) possible rules as presented in (1).Particularly, we first partition the \(i\)th input \({z}_{j}\) to overlapping membership functions (MFs), that are assumed to be normal, complete, and consistent. Then, using the data, normal, complete, and consistent fuzzy rules are generated adaptively in the form of (1),where rule index is \({\text{i}} \, = \, 1, 2, \ldots, {\text{R}}\), input variables are \({\text{z}}_{\text{j}} \, \left({\text{j}} \, = \, 1, 2, \ldots, {\text{m}}\right)\) and \(k\) is the output variable. The \({A}_{j}^{i}\) and \({B}^{i}\) are linguistic terms that are defined by inputs and output MFs of \({\mu }_{{A}_{j}^{i}}\left({z}_{j}\right)\) and \({\mu }_{{B}^{i}}\left(k\right)\), respectively.$${R}_{i}=\text{IF} {z}_{1} \,\text{is}\, {A}_{1}^{i} \,\text{and}\, {z}_{2} \,\text{is}\, {A}_{2}^{i} \,\text{and}\dots \text{and}\, {z}_{j} \,\text{is}\, {A}_{j}^{i} \,\text{then}\, k \,\text{is}\, {B}^{i},$$
(1)
We previously in Ref.7 proved that using triangular MFs, the sum of the multiplication of FBFs is equivalent to a common fuzzy system composed of a singleton fuzzifier, centroid defuzzifier, and product inference engine. Here, applying the proof, we use (2) for designing our fuzzy system.$$f\left(z\right)=\sum_{i=1}^{R}{d}_{i}\left(z\right){\theta }_{i},$$
(2)
where \({\theta }_{i}\) is the point in the output space at which the output membership function of \({\mu }_{{B}^{i}}\left(k\right)\) reaches its maximum value of 1 and \({d}_{i}\left(z\right)\) is the FBF as defined in (3).$${d}_{i}\left(z\right)=\prod_{j=1}^{m}{\mu }_{{A}_{j}^{i}}\left({z}_{j}\right),$$
(3)
where \({\mu }_{{A}_{j}^{i}}\left({z}_{j}\right)\) is input MF.
Nano fuzzy alarming system for blood transfusion requirement detection in cancer using deep learning
