IAES Inter national J our nal of Articial Intelligence (IJ-AI) V ol. 15, No. 4, August 2026, pp. 3944 ∼ 3954 ISSN: 2252-8938, DOI: 10.11591/ijai.v15.i4.pp3944-3954 ❒ 3944 Intra v enous immunoglob ulin r esistance pr ediction in Kawasaki disease using o v ersampled transf ormer embeddings Namitha Thattarassery Nanappan 1,2 , Ragha v endra Srini v asaiah 3 , V inith Rejathalal 4 1 Department of Computer Science and Engineering, CHRIST (Deemed to be Uni v ersity), Beng aluru, India 2 Department of Computer Science and Engineering, Mahaguru Institute of T echnology , Kayamkulam, India 3 Department of AI and Data Science Engineering, CHRIST (Deemed to be Uni v ersity), Beng aluru, India 4 Department of Articial Intelligence, Amrita V ishw a V idyapeetham, Coimbatore, India Article Inf o Article history: Recei v ed Jun 20, 2025 Re vised Jul 11, 2026 Accepted Jul 21, 2026 K eyw ords: Intra v enous immunoglob ulin resistance Ka w asaki disease Machine learning Ov ersampling T ransformer embeddings ABSTRA CT Ka w asaki disease (KD) is a leading cause of acquired heart disease in children under v e. Although intra v enous immunoglob ulin (IVIG) treatment is usually ef fecti v e, 10–20% of cases are resistant and at higher risk for coronary complications. Early prediction of IVIG resistance is critical b ut dif cult due to the rarity of KD and imbalanced clinic al data. T o address this, we propose a no v el technique called sentence transformer embeddings with synthetic minority o v er -sampling technique (SMO TE) o v ersampling (STESO), whi ch le v erages the complementary strengths of tra nsformer -based representation learning and synthetic o v ersampling. Pretrained models such as paraphrase-MiniLM-L3-v2 are used to con v ert tab ular clinical data into dense te xt-based embeddings, capturing deeper semantic relationships across features. By coupling these rich embeddings with SMO TE, we balance class distrib utions directly in the semantic space, enabling tradi tional machine learning (ML) models to more ef fecti v ely detect minority (resistant) cases. This syner gy yielded substantial impro v ements in sensiti vity and F1-score, with random forest (RF) combined with STESO (RF-STESO) achie ving the highest o v erall performance. Among the models e v aluated, our proposed model attained best result as accurac y of 0.85, sensi ti vity of 0.81, speci city of 0.89, and F1-score of 0.85. Our results underscore that the joint use of transformer embeddings and o v ersampling is more ef fecti v e than either approach in isolation, of fering a promising pathw ay for rare disease prediction tasks such as IVIG resistance prediction in KD. This is an open access article under the CC BY -SA license . Corresponding A uthor: Namitha Thattarassery Nanappan Department of Computer Science and Engineering, CHRIST (Deemed to be Uni v ersity) Beng aluru, Karnataka, India Email:namitha.tn@res.christuni v ersity .in 1. INTR ODUCTION Ka w asaki disease (KD), rst described by T omisaku Ka w asaki in 1967 and published in 1974, is a leading cause of acquired heart disease in children w orldwide. Children under the age of v e are the most af fected. The y e xperi ence prolonged fe v er , conjuncti v al injection, sw ollen mucosa, sw ollen lymph nodes, a characteristic rash, and mucosal Inammation. The cause of KD continues to be unanswerable, and the absence of treatment for KD can and will lead to se v ere long-term cons equences. Coronary artery aneurysm (CAA), J ournal homepage: http://ijai.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 3945 myocarditis, and arrh ythmias, to name a fe w , are all complications that can be de v eloped and/or e xacerbated if KD is not treated [1]. Japan has the highest incident rate of KD, with about 12,000 cases each year . The primary KD treatment has been and continues to be intra v enous immunoglob ulin (IVIG), and the coronary complications will be reduced with this treatment. Upon treatment, some patients demonstrate persistent or recurring fe v er 36 hours after the infusion. These patients at this point demonstrate IVIG treatment resistance. Risk o f coronary complications for these patients is around 9 times more lik ely than for those that respond to treatment [2]. The main treatment of KD is IVIG, and the most complicated portion of the treatment is predicting treatment resistance. The unbalanced datasets for KD ha v e o v erwhelmingly more responsi v e cases than cases resistant to treatment. Early identication of IVIG resistance is critical to pro vide additional interv entions, such as corticosteroids or iniximab, which can signicantly impro v e patient outcomes [3]. Predicting IVIG resistance remains challenging due to the imbalanced nature of KD datasets, where IVIG-responsi v e cases dominate. Sensiti vity scores of e xisting systems such as K obayashi, Eg ami, and Sano scores f all under 50%, compromising these systems o v erall predicti v e capability [4]. Ongoing research ef forts in this area are summarized in [5] and de v eloped a kno wledge frame w ork re g arding IVIG resistance predictions. Mirata et al . [6] put together a re vie w which suggested that a lar ge portion of present-day research focused on classical machine learning (ML) methods. There are classical ML techniques whic h include logistic re gression (LR), random forest (RF), support v ector machines (SVM), gradient boosting (GB), light gradient boosting machine (LightGBM), e xtreme gradient boosting (XGBoost), and neural n e tw orks (NN). F or e xample, Lam et al. [1] performed an e v aluation of the models which included LR, RF , nai v e Bayes, gradient boosting m achine (GBM), and NN and documented f airly lo w sensiti vity with LR scoring 0.216, and SVM 0.60. Sunag a et al . [7] sho wed implementation of LightGBM and documented sensiti vity of 0.60. The study , [8] incorporated LR, while [9] utilized a wide v ariety of models that consisted of LR, multi-layer perceptron (MLP), RF , cate gorical boosting (CatBoost), e xplainable boosting machine (EBM), and GBM. The papers [10] and [11] both directed their ef forts on models that use RF . In [12] XGBoost model sho wed strong performance on the testing set where the y scored an area under the curv e (A UC) of 0.821, an accurac y of 74.8%, sensiti vity of 88.9%, and specicity of 68.3%. Also, Liu et al. [13] used LR, SVM, XGBoost, and LightGBM. Ev en with the range of models and the gro wing v olume of datasets, most of the approaches continue to ha v e poor sensiti vity which is a serious problem gi v en the class imbalanc e present in KD datasets. Although performance metrics such as accurac y , specicity , and A UC are commonly reported, enhancing sensiti vity remains essential for making these prediction models clinically meaningful. This study inte grates ML with transfer learning and inno v ati v e natural language proces sing technologies to nd a solution. In clinical natural language processing, transformers are becoming essential technologies and are particularly and highly v ersatile at understanding dense and intricate medical te xts. Most recent research uses domain-specic models that are trained on healthcare materials and surpass general-purpose models for diagnosis coding and symptom e xtraction. F or instance, NorDeClin-bidirectional encoder representations from transformers (BER T) trained on Norwe gian clinical notes and signicantly enhanced results on international classication of diseases (ICD) classication [14]. Spanish transformers, for instance, enhanced e xtraction of heart f ailure symptoms from elec tronic health records (EHRs) [15]. It is no longer an academic e x ercise to ne-tune generati v e pre-trained transformers (GPTs) on medical corpora for clinical summarization and diagnos is support, sharpening the focus on the clinical specialization of healthcare natural language processing [16]. Unlik e the traditional hand-coded models, wherein numerical data are represented as te xts, language models , and more particularly sentence transformers, are used to nd the comple x te xts that describe the relationship between the data. Class imbalance is corrected using the synthetic minority o v er -sampling technique (SMO TE), and the enhanced embeddings are passed to the classical ML methods. This approach bridges the g ap between con v entional scoring systems and modern art icial inte lligence (AI) models, resulting in signicant impro v ements in prediction sensiti vity and o v erall performance. This w ork introduces sentence transformer embeddings with SMO TE o v ersampling (STESO) algorithm as a no v el frame w ork for early and accura te identication of IVIG-resistant KD patients. By situating STESO within the broader paradigm of clinical decision support systems (CDSS), this study sho ws ho w sentence-transformer embeddings can repurpose structured clinical data for sensiti v e prediction tasks, while laying the groundw ork for future inte gration with narrati v e records. Intr avenous immuno glob ulin r esistance pr ediction in Kawasaki disease ... (Namitha Thattar assery Nanappan) Evaluation Warning : The document was created with Spire.PDF for Python.
3946 ❒ ISSN: 2252-8938 2. D A T A AND METHODS The dataset used in this study , as described in [13], consis ts of 1,398 medical records of KD patients co v ering the years 2015-2020. The records include 1,240 cases of IVIG-responsi v e patients and 158 cases of IVIG-resistant patients, and contained 31 v ariables of demographic, clini cal, and imaging, and laboratory data. T o tackle the issue of data imbalance and to concentrate on impactful predictors, as described in the original study , the dataset underwent feature selection to determine rele v ant predictors using uni v ariate analysis via the chi-square test in SPSS v ersion 25.0 (p-v alue 0.05) [13]. F or the purposes of ensuring consistenc y and generalizability of the approach, we used the dataset in the research [17]. Since no predened feature subset w as a v ailable for this dataset, we emplo yed Shaple y additi v e e xplanations (SHAP) analysis on the full set of v ariables to identify the top 14 features contrib uting most to IVIG resistance predictions. SHAP analysis highlighted features such as day of fe v er and platelet count (PL T), both established clinical mark ers in KD, thereby supporting the biological plausibility of our approach. T o mak e the dataset compatible with a language transformer model, numerical features were transformed into ordinal cate gorical v alues by analyzing central t endenc y , spread, and quartile metrics. F or instance, continuous v ariables were cate gorized into le v els, ensuring consistenc y across all selected features. The processed cate gorical data w as further con v erted into structured sentences suitable for the sentence transformer [18]. One record yielded the follo wing clinical summary . The patient had high days of illness. L ymphoc yte count w as lymphoc yte medium, and hemoglobin (HB) le v els were HB medium. Neutrophil percentage w as neutr v ery v ery high, with an neutrophil-to-lymphoc yte ratio (NLR) of neutr ultimate maximum. Glob ulin w as glob ulin high while alb umin w as alb umin minimum. Aspartate aminotransferase (AST) le v els were AST less, and serum sodium w as serum sodium v ery v ery high. T otal bilirubin (TBIL) w as TBIL medium, procalcitonin (P CT) w as PCT maximum, and alanine aminotransferase (AL T) w as AL T medium. PL T le v els were PL T less, and g amma-glutamyl transferase (GGT) w as GGT minimum. This transformation allo wed the model to handle the dataset ef fecti v ely while maintaining the inte grity of critical information. These cate gory thresholds were cross-check ed ag ainst both statistical distrib ution patterns and established clinical rele v ance (e.g., lo w PL Ts and prolonged fe v er are well-documented risk f actors for IVIG resistance), thereby ensuring that the semantic transformation preserv ed medical plausibility . These preprocessing steps ensured the dataset w as optimized for the transformer model and highlighted the importance of rob us t preprocessing in predicti v e modeling for rare diseases lik e KD. As sho wn in Figure 1, adv anced ML techniques are still underutilized in this eld. Our proposed method bridges this g ap by using transformers as the base model. The method section describes in detail the underlying models-BER T , sentence-BER T (SBER T), and sentence transformer ne-tuning (SetFit)—before turning to the introduction of our ne w model, STESO. This ordering allo ws for a well-founded comprehension of the progression of the proposed architecture. BER T is a transformer -based language model designed for v arious tasks, such as te xt classication. It uses a bidirectional encoder and is pre-trained using self-supervised learning tasks, including mask ed language modeling (MLM) and ne xt sentence prediction (NSP) [19]. These tasks allo w BER T to capture the structural and semantic intricacies of hum an language without requiring labeled data. The MLM task randomly masks 15% of tok ens in the input and predicts them based on the surrounding conte xt. The NSP task trains the model to determine whether a gi v en sentence follo ws another logically . Architecturally , BER T consists of stack ed transformer encoders. The base model (BER TB ASE) has 12 layers, 768 hidden dimensions, and 12 self-attention heads, with 110M parameters in total, while the lar ge model (BER TLARGE) has 24 layers and 340M parameters. Despite its ef fecti v eness in numerous natural language processing tasks, BER T embeddings underperform in pairwise sentence tasks, such as clustering and similarity detection. T o address this limitation, SBER T w as proposed. Sentence transformer (SBER T) e xtends BER T by using a Siamese or triplet netw ork architecture to handle pairwise or triplet-based tasks such as classication and similarity comparison [20]. The model ne-tunes BER T embeddings to generate sentence-le v el representations that are more ef fecti v e for pairwise and clustering tasks [21]. In its architecture, pairs or triplets of sentences are passed through the netw ork simultaneously , generating embeddings that can be directly used for tasks lik e semantic similarity or classication. In cases where there is a need for quick comparisons, SBER T embeddings come in handy , especially for do wnstream tasks featuring sentence representations. SetFit e xpands on the concept of SBER T by applying a tw o-phase training methodology for fe w shot lea rning [22]. F or some recent w ork, fe w shot Int J Artif Intell, V ol. 15, No. 4, August 2026: 3944–3954 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 3947 learning has been tar geted in the medical eld which has data constrained medical tasks specically in rare diseases. F or instance, 3D fe w-shot learning w as de v eloped for a model that classies rare knee injuries in magnetic resonance imaging (MRI) [23] which sho wcases generalization to lo w-resource imaging tasks, a challenge in the eld of medical imaging. LR RF GBM LightGBM SVM XGBoost MLP/NN nai v e Bayes CatBoost EBM 0 2 4 6 3 4 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 ML Models Mentions 2021 2023 2024 2025 Figure 1. Model mentions across years Furthermore, generati v e data augmentation has been sho wn to impro v e performance in diagnostic visual recognition of fe w-shot cases which impro v es the recognition of certain conditions that are more dif cult to diagnose [24]. In a similar design, lar ge language models were used with ontologies in a rare pediatric epilepsy case [25] which pro vided a fe w-shot e xtraction of information re g arding drug ef cac y which sho ws the enhanced clinical natural language processing task of semantic inte gration. In the rst phas e, the training of the model u s es a contrasti v e technique where sentence pairs are made into positi v e and ne g ati v e e xamples. This process signicantly e xpands t h e training dataset size, scaling it quadratically to O ( K 2 ) , where K is the original sam ple size. This phase ne-tunes the sentence transformer to ef fecti v ely distinguish between contrasting sentences. In the second phase, embeddings generated by the ne-tuned transformer are used to train a do wnstream classier with labeled data. This combination of contrasti v e learning and classication mak es SetFit particularly ef fecti v e in data-scarce en vironments, of fering a rob ust and ef cient foundation for generating embeddings tailored to specic tasks. Figure 2 illustrates the o v erall training and inference process of the SetFit model. Figure 2. SetFit training and inference block diagram [22] Intr avenous immuno glob ulin r esistance pr ediction in Kawasaki disease ... (Namitha Thattar assery Nanappan) Evaluation Warning : The document was created with Spire.PDF for Python.
3948 ❒ ISSN: 2252-8938 2.1. Pr oposed sentence transf ormer embeddings with SMO TE o v ersampling methodology Ev en though BER T and other pre-trained language transformer models create sentence embeddings, the y do not detect subtle dif ferences especially in imbalanced datasets. Here, the g ap is addressed by using SMO TE in the embedding space for the minority embeddings helps classiers w ork better . The STESO methodology combines sentence transformer embeddings, SMO TE o v ersampling, and ordinary ML classiers to manage the e xtreme imbalance in class distrib ution. T o align with transformer models, we con v erted numerical features to te xtual sentences. These sentences are then encoded and run through the paraphrase-MiniLM-L3-v2 model from sentence transformers, and the embeddings are created by a v eraging the last hidden layer across all tok ens. This creates high-dimensional embeddings, or v ector r epresentations, e i , that capture v arious semantic and conte xtual aspects. These embeddings go be yond simple encoding by higher -order interaction modeling with SHAP-selected predictors, lik e characterizing prolonged fe v er and lo w PL T as a single clinical indicator . This allo ws the model to approximate clinician-style reasoning, impro v es generalization on imbalanced datasets, and maintains traceability to established predictors while enriching them with conte xtual meaning. Let the dataset be represented by (1). D = { ( x 1 , y 1 ) , ( x 2 , y 2 ) , . . . , ( x n , y n ) } (1) In (1), where each x i ∈ R 14 is a 14-dimensional numeric feature v ector , and y i ∈ { 0 , 1 } is the associated binary class label. Each feature v ector x i = [ x i 1 , x i 2 , . . . , x i 14 ] is transformed into a natural language sentence. This transformation is e xpressed through (2). S i = concat ( feature1 is x i 1 , feature2 is x i 2 , . . . , feature14 is x i 14 ) (2) As dened in (2), allo wing te xtual interpretation of numerical features. This sentence is passed through the pretrained sentence transformer model (STM) T Sent . The model tok enizes the input and encodes it into conte xtualized tok en-le v el embeddings are gi v en by (3). Encoder ( S i ) = [ h 1 , h 2 , . . . , h m ] (3) Where h j ∈ R d and m is the number of tok ens. These tok en embeddings are aggre g ated using mean pooling to obtain a x ed-size v ector as in (4). e i = 1 m m X j =1 h j , where e i ∈ R d (4) All embeddings are then stack ed to form the matrix. Stacking all embeddings forms the matrix in (5). E = e T 1 e T 2 . . . e T n ∈ R n × d (5) T o handle class imbalance, SMO TE is applied in the embedding space. The minority class embeddings are dened as (6). E + = { e i ∈ R d | y i = 1 } (6) Each e i ∈ E + is used to nd k nearest neighbors using Euclidean distance. The nearest neighbors are identied as in (7). NN ( e i ) = { e i 1 , e i 2 , . . . , e ik } ⊂ E + such that ∥ e i − e ij ∥ 2 is minimized (7) Ne w synthetic samples are generated through interpolation as sho wn in (8). e ne w = e i + λ · ( e j − e i ) , λ ∼ U (0 , 1) (8) Int J Artif Intell, V ol. 15, No. 4, August 2026: 3944–3954 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 3949 The resulting balanced dataset and corresponding labels are pro vided in (9) and (10). E SMO TE = E ∪ { e 1 ne w , e 2 ne w , . . . , e m ne w } (9) y SMO TE = y ∪ { 1 , 1 , . . . , 1 | {z } m } (10) The embeddings E SMO TE and labels y SMO TE can no w be used to train classical ML models. If LR is chosen, the predicted probabilities ˆ y i ∈ [0 , 1] are optimized using the binary c ross-entrop y loss. The STESO architecture is detailed in Algorithm 1. This study tak es adv antage of the ability of pretr ained sentence transformers to generate meaningful sentence embeddings from minimal data. These embeddings are then balanced using SMO TE to address clas s imbalance. By combining this wit h traditional classiers, STESO of fers a practical solution for achie ving strong performance e v en when data is limited. Algorithm 1 STESO: sentence transformer embeddings with SMO TE o v ersampling Requir e: tab ular dataset D = { ( x i , y i ) } n i =1 , where x i ∈ R 14 , y i ∈ { 0 , 1 } ; pretrained sentence transformer T sent ; SMO TE function. Ensur e: trained classier L and e v aluation metrics. 1: Sentence Con v ersion: for each x i , con v ert to clinical sentence S i 2: Sentence Embedding: e i = T sent ( S i ) 3: F orm embedding matrix: E = [ e 1 , ..., e n ] T ∈ R n × d 4: Ov ersampling train set: ( E ′ , y ′ ) = SMO TE ( E , y ) 5: T raining: train a classier L (e.g., RF , LR) on ( E ′ , y ′ ) 6: Ev aluation: compute accurac y , sensiti vity , specicity , and F1-score on test set 3. RESUL TS AND DISCUSSION This section compares traditional ML, deep learning (DL), SetFit-based fe w-shot learning with a STM, and the proposed STESO frame w ork. Results sho w progressi v e impro v ements across methods. On the unbalanced dataset of 1,398 patient re cords, traditional ML models (LR, RF , SVM, k-nearest neighbor (KNN)) and a feed-forw ard neural netw ork (FFNN) achie v ed high accurac y and specicity b ut performed poorly in predicting IVIG resistance, with lo w sensiti vity and F1-scores as in T able 1. T able 1. Performance metrics of con v entional ML and DL models using dataset [13] Metric LR RF SVM KNN FFNN STM Accurac y 0.92 0.93 0.90 0.90 0.86 0.88 Sensiti vity 0.36 0.48 0.15 0.24 0.42 0.48 Specicity 0.99 0.93 1.00 0.93 0..92 0.93 F1-score 0.53 0.70 0.41 0.37 0.46 0.47 T o e v aluate sentence transformer , this study ne-tuned t h e SetFit frame w ork using its pret rained sentence transformer backbone on balanced datasets with increasing shot sizes (25, 50, 100, 158). Sensiti vity impro v ed steadily , reaching 0.709 with 158-shot samples (Figure 3). Since lar ge datasets are uncommon in rare diseases such as KD, this study e xtended this idea by directly emplo ying the same STM to generate semantically rich embeddings from the full clinical dataset. Applying SMO TE in this embedding space produced balanced training data that preserv ed clinical structur e, enabling traditional ML models to more ef fecti v ely detect resistant cases. As sho wn in Figures 4 and 5, STESO consistently impro v ed sensiti vity and F1-scores, with RF achie ving the best o v erall performance due to its capacity to capture comple x non-linear interactions within the enriched feature space. These ndings underscore STESO’ s strength as a scalable solution for predicti v e modeling in rare and imbalanced medical datasets. Intr avenous immuno glob ulin r esistance pr ediction in Kawasaki disease ... (Namitha Thattar assery Nanappan) Evaluation Warning : The document was created with Spire.PDF for Python.
3950 ❒ ISSN: 2252-8938 LR RF SVM KNN FFNN STM 0 0 . 2 0 . 4 0 . 6 0 . 8 1 0 . 65 0 . 6 0 . 6 0 . 4 0 . 4 0 . 33 0 . 5 0 . 6 0 . 6 0 . 7 0 . 6 0 . 38 0 . 72 0 . 72 0 . 72 0 . 55 0 . 5 0 . 54 0 . 51 0 . 58 0 . 54 0 . 48 0 . 61 0 . 71 Model Sensiti vity 2-w ay 25 shot 2-w ay 50 shot 2-w ay 100 shot 2-w ay 158 shot Figure 3. Sensiti vity impro v ement of STM with increasing shot counts LR RF SVM KNN FFNN 0 0 . 5 1 Best Model Performance Metric Acc (Before) Acc (After) Sens (Before) Sens (After) Spec (Before) Spec (After) F1 (Before) F1 (After) Figure 4. Performance comparison of selected ML models before and after applying sentence transformer embeddings with SMO TE. RF-STESO sho ws the best o v erall performance across all metrics LR-Sen LR-F1 RF-Sen RF-F1 SVM-Sen SVM-F1 KNN-Sen KNN-F1 FFNN-Sen FFNN-F1 0 0 . 5 1 0 . 36 0 . 53 0 . 48 0 . 7 0 . 15 0 . 41 0 . 24 0 . 37 0 . 42 0 . 46 0 . 72 0 . 72 0 . 96 0 . 95 0 . 59 0 . 66 0 . 99 0 . 82 0 . 79 0 . 78 Model-Metric Performance Metric ML ML STESO Figure 5. Impro v ement in sensiti vity and F1-score with ML STESO T o e v aluate rob ustness be yond the initial dataset, this study applied STESO to a second highly imbalanced dataset [17]. As sho wn in T able 2, STESO impro v ed performance across all models (LR, RF , SVM, KNN, and FFNN). These g ains stem from the syner gy of transformer -based embeddings, which capture meaningful clinical patterns, and SMO TE, which balances minority cases in the semantic space. This combination enabled better detection of resistant cases compared to baselines. All impro v ements were conrmed as statis tically signicant using McNemar’ s test ( p < 0 . 05 ) with 10 fold cross v alidation with RF and RF-STESO, demonstrating STESO’ s rob ustness and adaptability for imbalanced datasets. Int J Artif Intell, V ol. 15, No. 4, August 2026: 3944–3954 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 3951 T able 2. Sensiti vity and F1-score impro v ements on dataset 2 [17] (ML vs ML STESO) to ensure model consistenc y and generalizability Model Sensiti vity (Before) Sensiti vity (After) F1-score (Before) F1-score (After) LR 0.14 0.52 0.24 0.32 RF 0.14 0.81 0.24 0.85 SVM 0.00 0.56 0.00 0.35 KNN 0.10 0.60 0.17 0.31 FFNN 0.35 0.46 0.26 0.60 According to the T able 3, the comparison sho ws that RF-STESO is superior to both con v olut ional neural netw orks (CNNs) which use con v olutional lters to capture local feature interactions, and the feature tok enizer transformer (FT -T ransformer) which is an attention-based model for tab ular data, pro viding greater and more balanced v alues across all the metrics of accurac y , sensiti vity , specicity , and F1-score. Ev en though CNNs deli v er acceptable perf o r mance and FT -T ransformers yield disproportionate outcomes across the metrics, RF-STESO still sho ws remarkable consistenc y across both datasets. This result underscores the po wer of inte grating semantic embeddings with RF to achie v e dependable forecasting for RF-STESO in clinically imbalanced scenarios. T able 3. Performance comparison of CNN, FT -T ransformer , and RF STESO across datasets Model/Dataset Accurac y Sensiti vity Specicity F1-score CNN [13] 0.89 0.83 0.83 0.83 FT -T ransformer[13] 0.90 0.27 1.00 0.43 RF STESO [13] 0.95 0.96 0.95 0.96 CNN [17] 0.83 0.93 0.74 0.82 FT -T ransformer [17] 0.80 0.32 0.94 0.41 RF-STESO [17] 0.85 0.81 0.89 0.85 3.1. Ablation study , failur e analysis, and futur e dir ections T o e v aluate each component of embedding generation and o v ersampling, in this ablation study tested v e scenarios: i) ML models trained on ra w clinical data, ii) ML models trained on ra w clinical data after SMO TE o v ersampling, iii) ML models trained on sentence-transformer embeddings with SMO TE o v ersampling, which is our proposed STESO frame w ork, i v) a comparati v e benchmark of STESO ag ainst DL approaches, CNN and FT -T ransformer , which w as trained directly on tab ular data as sho wn in T ables 3 and 4, and v) o v ersampled FT T ransformer embedding with sentence transformer embeddings as sho wn T able 5. Performance e v aluation sho wed that STESO resulted in the highest sensiti vity and F1-score. This illustrates the syner gistic ef fect that learning semantic representations and adjusting class distrib utions on detecting the minority class in IVIG resistance prediction. This model is a no v el decision-support conceptual tool that helps clinicians in the early identica tion of IVIG resistance. Such a model w ould enable clinicians to identify high-risk patients that require close monitoring, timely echocardiographic e v aluation, or the consideration of adjunct therapies. Ho we v er it’ s use is constr ained by l imitations such as the relati v ely small dataset, site-specic bias, and the need for e xternal v alidation before clinical deplo yment. In proposed method, some models sho wed a decline in accurac y and specicity . But in our case, the accurate prediction of IVIG resistance is more critical than predicting responsi v e patients. So, we mainly focused on impro ving the sensiti vity and F1-score. Other o v ersampling approaches lik e adapti v e synthetic (AD ASYN) also not e xplored in current study . Ev enthough, the embeddings enable the representation learning, these latent features are dif cult to interpret. T o impro v e interpretation, e xplainable techniqus lik e SHAP can be utilized in future. Future w ork will also include e xternal v alidation on geographically distinct cohorts to enhance generalizability , alongside ef forts to impro v e interpretability of transformer embeddings, as well as e xploring recent adv ances in multimodal fusion and graph-based modeling to capture richer clinical representations. While embeddings bridge structured and unstructured cli nical data for CDSS, ethical and deplo yment considerations remain critical. This study plan to address bias mitig ati on , strengthen data pri v ac y safe guards, conduct multi-center clinical v alidation, and e xplore inte gration into EHR systems and mobile diagnostics to support real-w orld adoption. Intr avenous immuno glob ulin r esistance pr ediction in Kawasaki disease ... (Namitha Thattar assery Nanappan) Evaluation Warning : The document was created with Spire.PDF for Python.
3952 ❒ ISSN: 2252-8938 T able 4. Comparison of LR, RF , SVM, and KNN under dif ferent data balancing strate gies using data [17] Method LR RF SVM KNN Sensiti vity ML models (no SMO TE) 0.14 0.14 0.00 0.10 ML + SMO TE (ra w features) 0.67 0.25 0.39 0.50 Sentence transformer embeddings + SMO TE + ML 0.52 0.81 0.56 0.60 F1-score ML models (no SMO TE) 0.24 0.24 0.00 0.17 ML + SMO TE (ra w features) 0.44 0.29 0.34 0.33 Sentence transformer embeddings + SMO TE + ML 0.32 0.85 0.35 0.31 T able 5. Comparison of sentence transformer and FT -T ransformer embeddings with SMO TE for selected ML models using data [17] Model Sensiti vity Sensiti vity F1-score F1-score (Sentence transformer) (FT -T ransformer) (Sentence transformer) (FT -T ransformer) LR 0.52 0.25 0.32 0.30 RF 0.81 0.25 0.85 0.36 SVM 0.56 0.37 0.35 0.42 KNN 0.60 0.37 0.31 0.37 4. CONCLUSION STESO demonstrates that combining sentence transformer -based embeddings with o v ersampling in the semantic space pro vides a po werful solution for IVIG resistance prediction in KD. By le v eraging pretrained sentence transformers, the proposed method in this study captures deeper clinical relationships, and applying o v ersampling in this enriched representation space generates synthetic resistant cases t h a t preserv e semantic and biological plausibility , unlik e ra w feature o v ersampling which can produce unrealistic combinations. This syner gy enhances sensiti vity and F1-score across models, with RF–STESO achie ving the strongest performance by e xploiting comple x interactions within the embeddings. Moreo v er , the embedding-space o v ersampling introduces controlled perturbations akin to re gularization, impro ving rob us tness and generalization to unseen data. Be yond KD, this strate gy of fers a transferable pathw ay for other rare disease prediction tasks where limited and imbalanced datasets remain a critical challenge, making transformer–o v ersampling inte gration a promising step to w ard more reliable and personalized healthcare solutions. FUNDING INFORMA TION Authors state there is no funding in v olv ed. A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the Contrib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib utions, reduce authorship disputes, and f acilitate collaboration. Name of A uthor C M So V a F o I R D O E V i Su P Fu Namitha Thattarasse y Nanappan ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Ragha v endra Srini v asaiah ✓ ✓ ✓ ✓ ✓ ✓ ✓ V inith Rejathalal ✓ ✓ ✓ ✓ ✓ ✓ ✓ C : C onceptualization I : I n v estig ation V i : V i sualization M : M ethodology R : R esources Su : Su pervision So : So ftw are D : D ata Curation P : P roject Administration V a : V a lidation O : Writing - O riginal Draft Fu : Fu nding Acquisition F o : F o rmal Analysis E : Writing - Re vie w & E diting Int J Artif Intell, V ol. 15, No. 4, August 2026: 3944–3954 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 ❒ 3953 CONFLICT OF INTEREST ST A TEMENT The author declares that there are no kno wn conicts of interest associated with this publication. Ther e are no nancial or personal relationships that could inappropriately inuence or bias the content of this w ork. INFORMED CONSENT Not applicable. This study did not in v olv e human participants, human data, or an y per sonally identiable information. All data used were either publicly a v ailable, fully anon ymized, or deri v ed from non-human sources, and therefore no informed consent w as required from indi viduals. ETHICAL APPR O V AL Not applicable. This research did not in v olv e human subjects, human biological materials, or e xperimental procedures on animals. The w ork w as conducted solely on computational models, publicly a v ailable datasets, or non-sensiti v e data that did not re q ui re interv ention with li ving or g anisms. Therefore, ethical appro v al from an institutional re vie w board or animal ethics committee w as not necessary for this study . D A T A A V AILABILITY The dataset used in this study is publicly a v ailable and w as obtained from pre viously published research as cited in references [13] and [17]. No ne w data were generated for this w ork. REFERENCES [1] J. Y . Lam et al ., “Intra v enous immunoglob uli n resistance in Ka w asaki disease patients: prediction using clinical data, ” P ediatric Resear c h , v ol. 95, no. 3, pp. 692–697, Feb . 2024, doi: 10.1038/s41390-023-02519-z. [2] Y . W ang et al ., “De v elopment of an immunoinammatory indicator -related dynamic nomogram based on machine learning for the prediction of intra v enous immunoglob ulin-resistant Ka w asaki disease patients, ” International Immunopharmacolo gy , v ol. 134, Jun. 2024, doi: 10.1016/j.intimp.2024.112194. [3] X. W ang, X. Shi, X. Guo, S. Chen, X. Lin, and F . Y ang, “Ef fecti v eness of initial corticosteroid treatment in Ka w asaki disease children suspected to be IVIG resistant, ” P ediatric Car diolo gy , v ol. 46, no. 8, pp. 2315–2321, Dec. 2025, doi: 10.1007/s00246-024-03657-9. [4] U. K. Akca et al ., “Comparison of IVIG resistance predicti v e models in Ka w as aki disease, ” P ediatric Resear c h , v ol. 91, no. 3, pp. 621–626, Feb . 2022, doi: 10.1038/s41390-021-01459-w . [5] J. Zhang et al ., “Kno wledge frame w ork of intra v enous immunoglob ulin resistance in the eld of Ka w asaki disease: a bibliometric analysis (1997-2023), ” Immunity , Inammation and Disease , v ol. 12, no. 5, May 2024, doi: 10.1002/iid3.1277. [6] D. Mirata et al ., “Learning-based models for predicting IVIG resistance and coronary artery lesions in Ka w asaki disease: a re vie w of technical aspects and study features, ” P ediatric Drugs , v ol. 27, no. 4, pp. 465–479, Jul. 2025, doi: 10.1007/s40272-025-00693-7. [7] Y . Sunag a et al ., “ A simple scoring model based on machine learning predicts intra v enous immunoglob ulin resistance in Ka w asaki disease, ” Clinical Rheumatolo gy , v ol. 42, no. 5, pp. 1351–1361, May 2023, doi: 10.1007/s10067-023-06502-1. [8] S. W ang et al ., “Establishment and v alidation of risk prediction model to predict intra v enous immunoglob ulin-resistance in Ka w asaki disease based on meta-analysis of 15 cohorts, ” Italian J ournal of P ediatrics , v ol. 51, no. 1, Feb . 2025, doi: 10.1186/s13052-025-01889-w . [9] E. J. Cheon, G. B. Kim, and S. P ark, “Predicti v e modeling of consecuti v e intra v enous immunoglob ulin treatment resistance in Ka w asaki disease: a nationwide study , ” Scientic Reports , v ol. 15, no. 1, Jan. 2025, doi: 10.1038/s41598-025-85394-4. [10] Y . Xia et al ., “ A machine learning-based model to predict intra v enous imm unoglob ulin resistance in Ka w asaki disease, ” iScience , v ol. 28, no. 3, Mar . 2025, doi: 10.1016/j.isci.2025.112004. [11] Y . He et al ., “Interpretable web-based machine learning model for predicting intra v enous immunoglob uli n resistance in Ka w asaki disease, ” Italian J ournal of P ediatrics , v ol. 51, no. 1, Jun. 2025, doi: 10.1186/s13052-025-02036-1. [12] L. Deng et al ., “Construction and v alidation of predicti v e models for intra v enous immunoglob ulin–resistant Ka w asaki disease using an interpretable machine learning approach, ” Clinical and Experimental P ediatrics , Jul. 2024, doi: 10.3345/cep.2024.00549. [13] J . Liu et al ., “ A machine learning model to predict intra v enous immunoglob ulin-resistant Ka w asaki disease patients: a retrospecti v e study based on the chongqing population, ” F r ontier s in P ediatrics , v ol. 9, No v . 2021, doi: 10.3389/fped.2021.756095. [14] P . D. Ngo et al ., “Domain-specic pretraining of NorDeClin-bidirectional encoder representations from transformers for international statistical classication of diseases, tenth re vision, code prediction in Norwe gian clinical te xts: model de v elopment and e v aluation study , ” JMIR AI , v ol. 4, pp. e66153–e66153, Aug. 2025, doi: 10.2196/66153. [15] J . Mata, V . P ach ´ on, A. Mano v el, M. J. Ma ˜ na, and M. D. L. V illa, “Multicriteria optimization of language models for heart f ailure with preserv ed ejection fraction symptom detection in Spanish electronic health records: comparati v e model ing study , ” J ournal of Medical Internet Resear c h , v ol. 27, Jul. 2025, doi: 10.2196/76433. [16] Y . H. Lim, P . S. Q. Y eoh, and K. W . Lai, “Ev aluating ne-tuned GPT models on dif ferent datasets in the healthcare domain, ” Inno vation and Emer ging T ec hnolo gies , v ol. 12, Jan. 2025, doi: 10.1142/S2737599425500124. [17] T . W ang, G. Liu, and H. Lin, “ A machine learning approach to predict intra v enous immunoglob ulin resistance in Ka w asaki disease patients: a study based on a Southeast Chi na population, ” PLoS ONE , v ol. 15, no. 8, Aug. 2020, doi: 10.1371/journal.pone.0237321. [18] L. T unstall et al ., “Ef cient fe w-shot learning without prompts, ” 2022, arXiv: 2209.11055 . Intr avenous immuno glob ulin r esistance pr ediction in Kawasaki disease ... (Namitha Thattar assery Nanappan) Evaluation Warning : The document was created with Spire.PDF for Python.