Indonesian J our nal of Electrical Engineering and Computer Science V ol. 42, No. 1, April 2026, pp. 215 ∼ 224 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v42.i1.pp215-224 ❒ 215 Multi-model deep ensemble framew ork f or early diagnosis of rar e genetic disorders using genomic, Phenotypic, and EHR data fusion Shan Mahmood 1 , Sayma Akter T rina 1 , Ar pita Saha Sukanna 1 , Sabrina Zaman Esha 1 , Md. Agdam Amin Adib 1 , Md. Sanim Ahmed 1 , Amirul Islam 2 1 Department of Computer Science and Engineering, American International Uni v ersity-Bangladesh, Dhaka, Bangladesh 2 Department of Electrical and Electronic Engineering, BSRM School of Engineering, BRA C Uni v ersity , Dhaka, Bangladesh Article Inf o Article history: Recei v ed Aug 9, 2025 Re vised Dec 13, 2025 Accepted Mar 4, 2026 K eyw ords: Deep learning Genetic disorder Healthcare Hybrid model Machine learning ABSTRA CT Rare genetic disorders pose signicant challenges in diagnosis because of their lo w pre v alence, heterogeneous manifestations, and lack of readily a v ailable datasets. This study systematically assesses v arious supervis ed and unsuper - vised m achine learning methods for the early diagnosis of rare genetic disorders based on a multi-center pediatric dataset of 2,434 anon ymized records enriched with demographic, clinical, and laboratory v ariables. In this study , genomic, phenotypic, and EHR v ariables were inte grated into a unied feature matrix, al- lo wing all modalities to be jointly analyzed within each m achine learning (ML) model. F ollo wing rigorous pre-processing steps, including the discard of nonin- formati v e identiers, imputation and encoding of cate gorical features, and nor - malization of numerical predictors, v e classication frame w orks were imple- mented: logistic re gression (LR), random forest (RF), one-dime nsional con v o- lutional neural netw ork (CNN), a h ybrid CNN long short-term memory (LSTM) model, and a stack ed ensemble of RF and XGBoost. Model performances were e v aluated on an independent test s et via accurac y , precision, recall, and F1-score metrics. While LR and the CNN baseline achie v ed F1-scores of 0.9090 and 0.8572, respecti v ely , tree-based models substantially outperformed deep learn- ing (DL) models: RF achie v ed an F1-score of 0.9565, and the CNN+LSTM h ybrid achie v ed 0.9611. RF+XGB ensemble achie v ed the highest diagnostic accurac y (98.77%) with balanced precision (0.9879) and recall (0.9877), illus- trating its superior capacity in capturing complicated, non-linear feature interac- tions and ghting ag ainst data imbalance. The results illustrate that bagging and boosting algorithms in combination pro vide a strong and interpretable frame- w ork for ef cient pre-screening of rare genetic disorders. The use of these ensemble techniques has the potential to enhance clinical practice by agging high-risk cases for v erication and f acilitating early therapeutic interv ention. This is an open access article under the CC BY -SA license . Corresponding A uthor: Amirul Islam Department of Electrical and Electronic Engineering, BSRM School of Engineering, BRA C Uni v ersity Dhaka, Bangladesh Email: amirul.islam@bracu.ac.bd J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
216 ❒ ISSN: 2502-4752 1. INTR ODUCTION Rare genetic diseases typically af fect fe wer than 4 to 5 indi viduals in e v ery 10,000. Y et collecti v ely , the y form a substantial w orldwide problem, inuencing in e xcess of 400 million indi viduals and demonstrating a combined pre v alence of 3.5 to 5.9 percent w orldwide. As much as 80 percent of them are genetic. Although there is no uniform international criterion, RDs are usually dened as those af fecting fe wer than 4–5 cases out of 10,000 indi viduals [1]. Considering them as a whole, RDs can be re g arded as a common e v ent, with 7,265, with an estimated accumulated pre v alence of 3.5–5.9% and af fecting more than 400 million people w orldwide [2]. Most RDs appear to be caused or modied by genetic f actors; u p to 80% of them are thought to ha v e a genetic etiology [3]. This points to the signicant necessity for rapid and precise diagnosis so that preliminary treatments, accurate genetic counseling, and impro v ed patient care can be addressed. In spite of progress in gene testing and medical diagnosis, achie ving a rm diagnosis is v ery challenging. Di v erse symptoms and the infrequent incidence of some syndromes lead to protracted diagnostic odysse ys, a high rate of misdiagnoses, and postponed treatment. Con v entional methods using sequential biochemical assays, single-gene tests, and specialist opinion generally do not ha v e the capacity to tackle numerous cases, recognize issues ef fecti v ely , or act suf ciently f ast to decipher complicated gene-symptom correlations on a grand scale. Also, the absence of lar ge, well-labeled classes from multiple centers and the huge class size v ariation mak e con v entional analysis methods dif cult. Ev en though rare genetic diseases are not common one by one, together the y af fect a lot of people around the w orld. These diseases are hard to diagnose because man y doctors do not ha v e much e xperience with them, and there is not al w ays enough data. Machine learning (ML) is a type of computer program that helps doctors understand health problems better . ML is a smart computer tool that can spot patterns in a person’ s genes and symptoms. It helps doctors nd out what illness someone might ha v e more quickly and accurately . One good e xample is DeepGestalt. It looks at f aces using deep learning (DL) to nd signs of o v er 215 genetic conditions. It gets the right answer in the top 10 guesses about 91% of the time. In some cases, it is done better than doctors [4]. Another tool is AlphaMissense, made by DeepMind. It checks small changes in DN A called missense mutations. W ith about 90% accurac y , it helps scientists gure out which changes might cause disease, so, the y can focus on the most important ones [5]. There is also SHEPHERD, from the Zitnik Lab . It uses patient data and DL to nd genes that might be causing a disease . It also matches patients with similar cases. This tool has helped a lot in the undiagnosed disease s netw ork [6]. Since there often is not enough labeled data in rare disease research, other learning methods are used. Sun and his team created a system that mix es unsupervised learning with techniques lik e self-distillation and gi ving the model guessed labels. It is useful, especially for diagnosing diseases from images [7]. Li et al. [8] used a type of model called a generati v e adv ersarial netw ork (GAN), which lets computers learn from lots of unlabeled data. Their method w ork ed better than re gular ones and sho wed that GANs are great for detecting rare diseases. Recently , researchers ha v e started combining dif ferent kinds of data to mak e models more accurate. F or e xample, W u and his team made Gestalt MML, which uses a T ransformer model to bring together f acial pictures, patient info, and doctor notes. This helps the system notice both visible and hidden symptom s [9]. Another tool is F ace2Gene from FDN A. T able 1 is the pre vious research on w orking in rare genetic dis orders, a model with performance (accurac y). Despite this signicant progress, prior literature on the detection of a rare genetic disorder still suf fers from se v eral k e y limitations: most deep-learning models, including DeepGestalt and GestaltMML, rely on lar ge curated image data, which is hard to generalize into f acial or phenotypic data- poor settings. Other methods, such as AlphaMissense and SHEPHERD, are po werful b ut narro wly focus on genomic v ariants and often miss important clinical and laboratory features that inform diagnosis. Impro v e- ments in lo w-label en vironments c o m e with semi-supervised and GAN-based approaches, most of which may yield unstable results and need careful tuning. Most i mportantly , v ery fe w studies ha v e combined these de v el- opments: genomic, phenotypic, and EHR data are inte grated within a single fused frame w ork, and cross-center v alidation is f ar too often lacking, limiting real-w orld clinical applicability . These g aps indicate that there is a great need for a unied, multi-modal, and rob ust diagnostic approach-an issue our study directly addresses. T o address these g aps, we pro vide a full ML pipeline applied to a uniform set of 2,434 anon ymous children’ s records. The records contain details on their background, health, and laboratory tests. Recent years ha v e seen a sur ge in interest in the application of articial intelligence (AI) and, in particular , ML algorithms because of their potential to re v eal intricate patterns in genetic data [10]. The accurac y of RD diagnosis has increased as a result of these ML algorithms’ demonstrated ability to learn from and act upon massi v e, di- v erse datasets in order to deri v e no v el biological insights [11], [12]. Examining the role of AI/ML algorithms Indonesian J Elec Eng and Comp Sci, V ol. 42, No. 1, April 2026: 215–224 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng and Comp Sci ISSN: 2502-4752 ❒ 217 in the diagnosis and prognosis of RDs using genomic data [3]. Genetic disorders result from abnormalities in DN A; each is usually rare, b ut tak en together , the y are a common cause of disease throughout the w orld. Symptoms are v aried, often o v erlapping, and clinical diagnosis is frequently v ery slo w . Early treatment is usu- ally essential for the best outcomes, yet traditional methodologies can be limited and sometimes inconclusi v e. Therefore, reliable data-dri v en models are ur gently needed to support f aster and more accurate identication of geneti c disorders. Our whole process encompasses thorough data preparation (remo v al of personal infor - mation, imputation of missing data, encoding labels, and normalization of data) and de v eloping v e methods for classifying the data. Among these, the random forest (RF)+Boost ensemble emer ged as the best-performed model, achie ving 98.77% accurac y and an F1-score of 0.9877 by ef fecti v ely capturing comple x, non-linear feature interactions and mitig ating class imbalance. Our study uniquely inte grates genomic, phenotypic, and EHR features into a single fused model and e v aluates v e traditional, DL, and ensemble approaches to identify the most reliable diagnostic frame w ork. The main contrib utions are as follo ws: − Computed and analyzed feature importance to pro vide meaningful insights for clinical practice, enabling early detection and interv ention of unusual genetic diseases. − Designed and implemented our proposed tw o model ensemble architectures (CNN+LSTM, RF+XGBOOSTER) to capture rare genetic disorders. The acceptability of the ensemble m od e l has been determined through v ar - ious indicators of accurac y , F1-score, precision, and recall. T able 1. Summary of rare disease detection models and their performance Ref. Model and method Data and task Reported performance [4] DeepGestalt: CNN-based f acial phe- notype frame w ork quantifying simi- larities to genetic syndromes 26,000+ patient cases across 215 syn- dromes; identify syndrome from uncon- strained 2D f acial images 91% T op-10 accurac y; outper - formed clinical e xperts in three e xperiments [5] AlphaMissense: Unsupervised lan- guage model ne-tuned wit h structural conte xt and e v olutionary conserv ation Proteome-wide missense v ariant pathogenicity pre diction across the hu- man proteome > 90% precision for kno wn clinical impact of v ariants [6] SHEPHERD: Fe w-shot DL o v er a biomedical kno wledge graph (dis- eases, phenotypes, genes) 465 real patients (299 dis eases) from the Undiagnosed Diseases Netw ork; tasks: causal gene disco v ery , “patients-lik e-me” retrie v al, phenotype characterization Causal genes rank ed at 3.52 on a v erage [7] Hybrid URL + Pseudo-Label Self- Distillation: Contrast i v e unsuper - vised representation learning inte- grated with pseudo-label supervised self-distillation Rare skin lesion classication on ISIC 2018 (fe w-shot setting with base dataset of com- mon diseases and controls) Substantially outperforms e x- isting fe w-shot learning meth- ods [8] Semi-supervised GAN (feature- matching + pull-a w ay term) for rare disease detection IQVIA longitudinal claims: 5,923 positi v es, 17,769 matched ne g ati v es, 1.17 M unla- beled (test: 23,246 positi v es of 1.77 M) 34.18% PR-A UC (vs. LR 29.04%, NN 28.95%, RF 10.51%) 2. METHOD In this methodology part, we present a clear e xplication of the data and step-by-step processes fol- lo wed in our study . First, we e xpound on the dataset used in the study in terms of its source, nature, and rele v ant features. W e then elaborate on the strong pre-processing processes and con v ert the data into a ML model-ready format. Secondly , we clarify the v arious supervised and h ybrid ML models used, describing their architectures. Finally , we specify the e v aluation to compare the performance of the implemented models. 2.1. Dataset description F or our project, we used “Genetic Disorder Dataset” from Kaggle. The data set is a retrospec ti v e, multi-center cohort of 2,434 anon ymized pediatric patient records (age range: 0–14 years; mean ± SD: 6.99 ± 4.38 years) from four tertiary care centers. Each record is assigned a unique, de-identi ed patient code and annotated with minimal demographic metadata (i nstitution name and location) to preserv e pro v enance without violating condentiality [13]. T o supplement these data, the data set includes quantitati v e lab tests, red and white blood cell counts e xpressed in nati v e units, and binary blood-test outcomes (normal, inconclusi v e, or missing represented as –99) [14]. Fi v e binary symptom ags record the occurrence or non-occurrence of primary clinical features, and the principal outcome measure “Genetic Disorder” (1 = present risk; 0 = not Multi-model deep ensemble fr ame work for early dia gnosis of r ar e g enetic ... (Shan Mahmood) Evaluation Warning : The document was created with Spire.PDF for Python.
218 ❒ ISSN: 2502-4752 present) is complemented by a free-te xt eld stating the cate gory of disorder . Figure 1 represents the la bel of our dataset, where 0 is no disorder , and 1 is disorder . Figure 1. Genetic disorder label 2.2. Dataset pr e-pr ocessing In our dataset, non-informati v e identiers were remo v ed, missing v alues were imputed, cate gorical v ariables were label-encoded, and numerical features were standar d i zed to prepare the datas et for modeling. These steps ensured a clean, consistent feature space suitable for all machine-learning models without altering the underlying clinical patterns. 2.3. Model In our paper , we applied three single ML and DL models and tw o h ybrid models to detect rare genetic disorders. Figure 2 represents all the models we applied in our paper , including our proposed model. Figure 2. Applied models o v ervie w 2.3.1. Logistic r egr ession The logistic re gression (LR) model is an open, baseline detector of rare genetic disorders. Input fea- tures, ha ving been preprocessed and encoded, are passed through a single dense layer that computes a weighted sum of all predictors [15]. This linear combination i s then passed through a logistic acti v ation function to output a probability score of the presence of a genetic anomaly . It is trained using maximum-lik elihood estimation with gradient-based optimization, L2 re gularization for coef cient size limiting, and o v ertting pre v ention [16]. Its computational tractability ensures rapid con v er gence, minimal memory usage, and reproducible performance in a wide v ariety of computing en vironments [17]. As a rst-line model, it of fers a performance benchmark ag ainst which more adv anced architectures can be rigorously compared. Indonesian J Elec Eng and Comp Sci, V ol. 42, No. 1, April 2026: 215–224 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng and Comp Sci ISSN: 2502-4752 ❒ 219 2.3.2. Random f or est A RF classier is a collection of decision trees. Each tree is trained on a bootstrap sample from the data, and trees can gro w to a predetermined maximum depth or until leaf-size constraints are met, balancing the trade-of f between bias and v ariance [18]. The nal prediction is obtained by majority v ote across all trees, and class-probability predictions are calculated by a v eraging indi vidual tree v otes. This model is disco v ering comple x, nonlinear interactions between clinical, genetic, and en vironmental v ariables without e xplicit feature engineering. P arallelization training and inference enable the RF to scale to lar ge pediatric cohorts [19]. RF model tak es adv antage of the ensemble of decision trees to spot nonlinear interactions among clinical and genetic v ariables. Gro w trees until either a minimum leaf-size or maximum depth threshold is reached to ensure di v ersity of the ensemble [20]. T rees cast v otes at inference 145 on the e xistence of a genetic disorder; such v otes are tallied by majority (or a v eraged for probability estimation) [21]. Its modular and intrinsic structure allo ws for easy scaling to lar ge cohorts through distrib uted tree construction [22]. 2.3.3. CNN CNN accepts each patient’ s record as a feature sequence so that local patterns can be e xtracted from neighboring v ariable sets. The model architectur e is composed of multiple con v olutional blocks, each with a con v olutional layer of small k ernel size, batch-normalization, ReLU acti v ation, and max-pooling [23]. A global a v erage-pooling layer then reduces each feature map to a scalar [24]. The nal sigmoid acti v ation produces the probability of a genetic disorder . W eight sharing and local connecti vi ty reduce the total parameter count, f acilitating generalization on moderate-sized clinical datasets [25]. The input tensor under goes a sequence of con v olutional blocks—each consisting of a one-dimensional con v olution (s mall k ernel), batch normalization, ReLU acti v ation, a nd max-pooling step-by-step learning hierarchical representations [26]. Finally , a dropout- re gularized fully connected layer computes the disorder probability with sigmoid acti v ation. This layering automatically learns comple x inter -feature relationships [26]. 2.3.4. Hybrid CNN+LSTM CNN+LSTM model inte grates con v olutional feat u r e learning and recurr ent sequence modeling to learn local patterns and long-range dependencies across feature windo ws. The early Con v1D blocks are anal- ogous to an independent CNN and pro vide a lo w-dimensional feature sequence. This is passed through an LSTM layer that has hidden states to capture information from all time steps. A dense output layer with sig- moid acti v ation pro vides the nal probability [27]. LSTM’ s g ating beha vior enables selecti v e memory of k e y features, enhancing the sensiti vity to atypical e v ent patterns. Empirical e xperi ments demonstrate this tw o-stage approach has a propensity to surpass entirely con v olutional or recurrent netw orks when it comes to e xtracting both sequence-le v el as well as motif-le v el information [28]. Con v olutional blocks (Con v1D BatchNorm ReLU MaxPool) [27] initially map continuous subsets of features into a lo wer -dimensional sequence. A nal dense layer with sigmoid acti v ation produces the probability estimate. Combining the con v olutional l- ters’ po wer and the LSTM’ s ability , the h ybrid is particularly ef fecti v e at identifying dif fuse characteristics of uncommon genetic disorders [29]. 2.3.5. Hybrid random f or est and gradient boosting F or enhancing the performance and stability of classication processes on our data, we propose an ensemble model combining RF and gradient boosting (GB) classiers with a soft v oting strate gy . Ensemble learning is a widely used approach to strengthen prediction capacity by combining the strengths of ensemble learners [30], [31]. In our method, RF and GB outputs are combined based on their estimated class probability , and the nal label is decided based on the a v eraged probabilities (soft v oting). RF is a collection of decision trees, and each tree casts a v ote for making the nal prediction. Its strengths are its rob ustness to o v ertting, its ability to learn non-linear relationships, and its ability to handle lar ge datasets. GB, on the other hand, sequentially b uilds learners, and e v ery ne w learner focuses on the errors of the e xisting one. It is reno wned for its good predicti v e performance and is susceptible to o v ertting and tuning parameters. Through the fusi on of the tw o models, we seek to le v erage their di v ersity and complementarity of learning paradigms: RF is kno wn to of fer stability and reduction of v ariance, whereas Gradient Boosting is aimed at bias correction and rened learning. Multi-model deep ensemble fr ame work for early dia gnosis of r ar e g enetic ... (Shan Mahmood) Evaluation Warning : The document was created with Spire.PDF for Python.
220 ❒ ISSN: 2502-4752 2.4. Ev aluation metrics In order to rigorously quantify and compare the diagnostic accurac y of each proposed classier , we apply the confusion-m atrix paradigm and four resultant summary measures, namely , confusion matrix, accu- rac y , precision, recall, and F1-score. Supporting our e v aluation is the confusion matrix, which holds model predictions ag ainst ground truth labels in a binary situation. It distinguishes between true positi v es (TP), f alse positi v es (FP), f alse ne g ati v es (FN), and true ne g ati v es (TN), thus illuminating whether errors result from f alse ne g ati v es [32]. Accurac y estimates the proportion of all correctly cl assied instances and is an intuiti v e es- timate of the o v erall correctness of the model. Precision is the fraction of correctly predicted disorder cases among predicted positi v es. High precision helps limit redundant follo w-up tests for f alse alarms. Recall esti- mates ho w well the model can pick actual instances of disorder from all the real positi v es [33]. The F1-score balances recall and precision into a scalar by their harmonic mean, yieldi ng a balance measure that is unique for class imbalance [33]. 3. RESUL TS AND DISCUSSION An e xtensi v e comparati v e study w as carried out to compare the performance of v e v aried m achine- learning setups in predicting rare genetic dis eases from intricate genomic and clinical datasets. The models in question were a linear LR classier , an ensemble bagged RF , a con v olutional neural netw ork (CNN), a CNN+LSTM netw ork, and a stack ed ensemble of RF with XGBoost (RF+XGB). Performance w as e v aluated o v er an independent test set, where accurac y , precision, recall, and F1-score were used as the primary metrics. LR achie v ed a baseline accurac y of 90.91% and an F1-score of 0.9090, reecting the inability of linear deci- sion boundaries to model the intricate, non-linear relationships inherent to rare disease genom ics. The CNN model, which w as created for local sequence motif identication, achie v ed a score of 85.71% (F1 = 0.8572), reecting its relati v e lack of ef fecti v eness when transferred to tab ular formats of genetic v ariants without sig- nicant domain-specic architectural modication or massi v e data augmentation. RF presented a dramatic impro v ement from LR and CNN with 95.65% accurac y and an F1-score of 0.9565. This dramatic boost is a testament to the ef cac y of decision tree ensembles at learning intricate feature interactions and mitig ating v ariance by bootstrap aggre g ating. The CNN+LSTM h ybrid architecture, which marries con v olutional lters for motif capture with recurrent layers for modeling sequence dependence, took it a step further with 96.10% accurac y and an F1-score of 0.9611. While the g ain o v er RF w as modest, it w as statistically signicant, indi- cating that the addition of ordered or sequential patterns, i.e., v ariant phasing or longitudinal clinical measures, yields additional predicti v e v alue. The best results were obtained by the RF+XGB ensemble that posted e xcellent metrics across the board: 98.77% accurac y , 0.9879 precision, 0.9877 recall, and an F1-score of 0.9877. These ndings represent a roughly 2.7-percentage-point impro v ement o v er CNN+LSTM and a 3.1-point impro v ement o v er RF alone, indicating the ensemble’ s better discriminati v e po wer in the rare-disease setting. T able 2 sho ws the applied algorithm and its performance matrix (accurac y , preci sion, recall, F1-score). Here, RF+XGB achie v ed a better result than other algorithms. T able 2. Performance comparison of dif ferent algorithms Algorithm Accurac y Precision Recall F1-score LR 0.9091 0.9093 0.9091 0.9090 RF 0.9565 0.9578 0.9565 0.9565 CNN 0.8571 0.8575 0.8571 0.8572 CNN + LSTM 0.9610 0.9614 0.9610 0.9611 Radom f or est + XGBoost (RF+XGB) 0.9877 0.9879 0.9877 0.9877 The ndings are important because the RF+XGB model presents v ery reliable performance for early rare-genetic-disorder detection, reaching an accurac y of 98.77 percent and a strong o v erall balance in precision and recall. The high performance here indicates that ensemble learning can underpin f aster and more accurate clinical screening. Additional genomic sequencing data could further this w ork, testing the model on lar ger multi-center datasets and using e xplainabl e AI tools to understand feature importance. K e y e xperiments that should be done include e xternal v alidation, ablation studies, a n d rob ustness testing under class imbalance. The main tak ea w ay is that the ensemble-based models pro vide a practical and po werful basis for impro ving diagnosis in early-stage rare diseases. Although the CNN+LSTM model slightly outperformed RF , tree-based Indonesian J Elec Eng and Comp Sci, V ol. 42, No. 1, April 2026: 215–224 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng and Comp Sci ISSN: 2502-4752 ❒ 221 methods, particularly the RF+XGB ensemble, still pro vided the strongest o v erall performance, indicating that ensemble strate gies capture nonlinear interactions more ef fecti v ely than single deep models. Figure 3 represents the point plot of performance metrics for all the algorithms, while Figure 4 sho ws the confusion matrix for RF+XGB. Logistic Regression Random Forest CNN CNN + LSTM RF + XGBoost Score 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Accuracy Precision Recall F1-Score Figure 3. Point plot of metrics by v arious algorithms Figure 4. Confusion matrix of proposed algorithm (RF+XGB) 4. CONCLUSION This research presents a holistic assess ment of v arious ML strate gies for early diagnosis of rare genet ic diseases by comparing con v entional linear models, deep-learning structures, and ensemble classiers ag ainst intricate genomic and clinical data. The ndings cate gorically indicate that ensemble tree methods are the best predictors with an accurac y rate of 98.77% and an F1-score of 0.9877, achie v ed by the RF+XGB model. The superior performance of RF+XGB model, in comparison to LR (accurac y: 90.91%), CNN (accurac y: 85.71%), and the h ybrid CNN+LSTM netw ork (accurac y: 96.10%). The signicant impro v ement pro vided by the RF+XGB ensemble is due to its tw o inherent strengths: RF v ariance-reducing bagging method and XGBoost’ s bias-reducing, re gularized gradient-boosting mechanism. The h ybrid model ef fecti v ely balances the risks of undertting and o v ertting. In addition, the ensemble interpretability is remarkable. In conclusion, our results demonstrate that the RF+XGB ensemble is a rob ust and interpretable basis for early diagnosis of Multi-model deep ensemble fr ame work for early dia gnosis of r ar e g enetic ... (Shan Mahmood) Evaluation Warning : The document was created with Spire.PDF for Python.
222 ❒ ISSN: 2502-4752 rare genetic disorders in complicated genomic and clinical data sets, of fering superior predicti v e reliability and strong potential for inte gration into modern intelligent healthcare and IoT -supported diagnostic systems. This w ork will also contrib ute to intelligent computing and healthcare IoT systems by pro viding a reliable data-dri v en diagnostic frame w ork. This ensemble model can be incorporated into smart clinical platforms for real-time screening and decision support. In general, the approach strengthens the link between ML, healthcare automation, and modern computational system design. FUNDING INFORMA TION This research w as funded by Shan Mahmood, Sabrina Zaman Esha, and Arpita Saha Sukanna, who acquired the nancial support necessary to carry out the study . A UTHOR CONTRIB UTIONS ST A TEMENT This journal uses the C o nt rib utor Roles T axonomy (CRediT) to recognize indi vidual author contrib u- tions, 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 Shan Mahmood ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Sayma Akter T rina ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ Arpita Saha Sukanna ✓ ✓ ✓ ✓ ✓ Sabrina Zaman Esha ✓ ✓ ✓ ✓ Md. Agdam Amin Adib ✓ ✓ Md. Sanim Ahmed ✓ Amirul Islam ✓ ✓ ✓ ✓ ✓ 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 and E diting CONFLICT OF INTEREST ST A TEMENT The authors state no conict of interest. D A T A A V AILABILITY Data supporting this study are a v ailable from the corresponding author upon request. REFERENCES [1] T . Richter et al. , “Rare disease terminology and denitions—a systematic global re vie w: report of the ispor rare disease special interest group, ” V alue in Health , v ol. 18, no. 6, pp. 906–914, Sep. 2015, doi: 10.1016/j.jv al.2015.05.008. [2] S. N. W akap et al. , “Estimating cumulati v e point pre v alence of rare diseases: analysis of the or -phanet database, ” European Journal of Human Genetics , v ol. 28, no. 2, pp. 165–173, 2020, doi: 10.1038/s41431-019-0508-0. [3] S. Brasil, C. P ascoal, R. Francisco, V . dos Reis Ferreira, P . A. V ideira, and G. V alad˜ao, “ Articial intelligence (AI) in rare diseases: is the future brighter?” Genes , v ol. 10, no. 12, p. 978, 2019, doi: 10.3390/genes10120978. [4] Y . Guro vich et al. , “Deepgestalt-identifying rare genetic syndromes using deep learning, ” arXi v preprint arXi v:1801.07637 , Jan. 2018, doi: 10.48550/arXi v .1801.07637. [5] J . Cheng et al. , “ Accurate proteome-wide misse nse v ariant ef fect prediction with alphamissense, ” Science , v ol. 381, no. 6664, p. eadg7492, Sep. 2023, doi: 10.1126/science.adg7492. [6] E. Alsentzer et al. , “Deep learning for diagnosing patients with rare genetic diseases, ” medRxi v , Dec. 2022, doi: 10.1038/s41746- 025-01749-1. [7] J . Sun, D. W ei, K. Ma, L. W ang, and Y . Zheng, “Unsupervised representation learning meets pseudolabel supervised self-distillation: A ne w approach to rare disease classication, ” in Proc. International Conference on Medical Image Computing and Computer - Assisted Interv ention , Strasbour g, France, Sep. 2021, pp. 519–529, doi: 10.48550/arXi v .2110.04558. [8] W . Li, Y . W ang, Y . Cai, C. Arnold, E. Zhao, and Y . Y uan, “Semi-supervised rare disease detection using generati v e adv ersarial netw ork, ” arXi v preprint arXi v:1812.00547 , Dec. 2018, doi: 10.48550/arXi v .1812.00547. Indonesian J Elec Eng and Comp Sci, V ol. 42, No. 1, April 2026: 215–224 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng and Comp Sci ISSN: 2502-4752 ❒ 223 [9] D. W u et al. , “Gestaltmml: Enhancing rare genetic disease diagnosis through multimodal machine learning combining f acial images and clinical te xts, ” ArXi v , pp. arXi v–2312 , Apr . 2024, doi: 10.48550/arXi v .2312.15320. [10] E. Routhier and J. Mozziconacci, “Genomics enters the deep learning era, ” PeerJ , v ol. 10, p. e13613, Jun. 2022, doi: 10.7717/peerj.13613. [11] J. Schaefer , M. Lehne, J. Schepers, F . Prasser , and S. Thun, “The use of machine learning in rare diseases: a scoping re vie w , ” Orphanet Journal of Rare Diseases , v ol. 15, no. 1, p. 145, Jun. 2020, doi: 10.1186/s13023-020-01424-6. [12] S. T . Setty , M.-P . Scott-Bo yer , T . Cuppens, and A. Droit, “Ne w de v elopments and possibilities in reanalysis and reinterpretation of whole e xome seque ncing datasets for unsolv ed rare diseases using machine learning approaches, ” International Journal of Molecular Sciences , v ol. 23, no. 12, p. 6792, Jun. 2022, doi: 10.3390/ijms23126792. [13] A. Raza te xtslet al., “Predicting genetic disorder and types of disorder using chain classier approach, ” Genes , v ol. 14, no. 1, p. 71, Dec. 2022, doi: 10.3390/genes14010071. [14] N. Chaplot , D. P ande y , Y . K umar , and P . S. Sisodia, “ A comprehensi v e analysis of artici al intelligence techniques for the prediction and prognosis of genetic disorders using v arious gene disorders, ” Archi v es of Computational Methods in Engineering , v ol. 30, no. 5, pp. 3301–3323, Jun. 2023, doi: 10.1007/s11831-023-09904-1. [15] S. Dom ınguez-Almendros, N. Ben ıtez-P arejo, and A. R. Gonzalez-Ramirez, “Logistic re gression models, ” Aller gologia Et Im- munopathologia , v ol. 39, no. 5, pp. 295–305, Sep. 2011, doi: 10.1016/j.aller .2011.05.002. [16] J. M. H ilbe, Logistic re gression models , Chapman and Hall/CRC, May 2009. [17] T . G. Nick and K . M. Campbell, Logistic re gression. T oto w a , NJ: Humana Press, Jan. 2007. [18] S. J. R ig atti, “Random forest, ” Journal of Insurance Medicine , v ol. 47, no. 1, pp. 31–39, 2017, doi: 10.17849/insm-47-01-31-39.1. [19] M. Belgiu and L. Dr agut ¸ , “Random forest in remote sensing: A re vie w of applications and future directions, ” ISPRS Journal of Photogrammetry and Remote Sensing , v ol. 114, pp. 24–31, Apr . 2016, doi: 10.1016/j.isprsjprs.2016.01.011. [20] A. P armar , R. Katariya, and V . P atel, “ A re vie w on random forest: An ensemble classier , ” in Proc. International Conference on In- telligent Data Communication T echnologies and Internet of Things , Coimbatore, India, Aug. 2018, pp. 758–763, doi: 10.1007/978- 3-030-03146-6 86. [21] L.-C. Chuang and P .-H. K uo, “Building a genetic risk model for bipolar disorder from genome-wide association data with random forest algorithm, ” Scientic Reports , v ol. 7, no. 1, p. 39943, Jan. 2017, doi: 10.1038/srep39943. [22] A. K. Kalusi v aling am, A. Sharma, N. P atel, and V . Singh, “Le v eraging deep learning and random forest algorithms for en- hanced genomic analysis in rare disease identication, ” International Journal of AI and ML , v ol. 2, no. 10, No v . 2013, doi: 10.1016/j.csbr .2025.100038. [23] F . Elmaz, R. Eyck erman, W . Casteels, S. Latr e, and P . Hellinckx, “CNN-LSTM architecture for predicti v e indoor temperature modeling, ” Building and En vironment , v ol. 206, p. 108327, Dec. 2021, doi: 10.1016/j.b uilden v .2021.108327. [24] K. Gupta, N. Jiw ani, and N. Afreen, “Blood pressure detection using CNN-LSTM model, ” in Proc. 2022 IEEE 11th Interna- tional Conference on Communication Systems and Netw ork T echnologies (CSNT) , Indore, India, Apr . 2022, pp. 262–366, doi: 10.1109/CSNT54456.2022.9787648. [25] Y . F an, H. Xiong, and G. Sun, “Deepasdpred: a CNN-LSTM-based deep learning method for autism spectrum disorders risk rna identication, ” BMC Bioinformatics , v ol. 24, no. 1, p. 261, Jun. 2023, doi: 10.1186/s12859-023-05378-x. [26] M . M. Hossain et al. , “Cardio v ascular disease identication using a h ybrid CNN-LSTM model with e xplainable AI, ” Informatics in Medicine Unlock ed , v ol. 42, p. 101370, Jan. 2023, doi: 10.1016/j.imu.2023.101370. [27] Z. V ujo vi c et al. , “Classication model e v aluation metrics, ” International Journal of Adv anced Computer Science and Applica- tions , v ol. 12, no. 6, pp. 599–606, Jan. 2021, doi: 10.14569/IJ A CSA.2021.0120670. [28] J . Liang, “Confusion matrix: machine learning, ” POGIL Acti vity Clearinghouse , v ol. 3, no. 4, Dec. 2022, doi: https://pac.pogil.or g/inde x.php/pac/article/vie w/304. [29] K. O’ shea and R. Nash, “ An introduction to con v olutional neural netw orks, ” arXi v preprint arXi v:1511.08458 , No v . 2015, doi: 10.48550/arXi v .1511.08458. [30] L. Rokach, “Ensemble-based classiers, ” Articial Intelligence Re vie w , v ol. 33, no. 1, pp. 1–39, 2010, doi: 10.1007/s10462-009- 9124-7. [31] P . Chujai, K. Chomboon, P . T eerarassamee, N. K erdprasop, and K. K erdprasop, “Ensemble learning for imbal anced data classi- cation problem, ” in Proc. of the 3rd International Conference on Industrial Application Engineering , v ol. 467, Kitak yushu, Japan, Mar . 2015, pp. 449–456, doi: 10.12792/iciae2015.079. [32] S. S. Y ada v and S. M. Jadha v , “Deep con v olutional neural netw ork based medical image classication for disease diagnosis, ” Journal of Big Data , v ol. 6, no. 1, pp. 1–18, Dec. 2019, doi: 10.1186/s40537-019-0276-2. [33] S. Sarraf and G. T oghi, “Classication of alzheimer’ s disease using fMRI data and deep learning con v olutional neural netw orks, ” arXi v preprint arXi v:1603.08631 , Mar . 2016, doi: 10.48550/arXi v .1603.08631 F ocus to learn more. BIOGRAPHIES OF A UTHORS Shan Mahmood is an Americ an International Uni v ersity-Bangladesh B.Sc. Computer Science and Engineering student. His areas of rese arch are AI, ML, DL, natural language processing (NLP), generati v e adv ersarial netw orks, and high-accurac y AI-dri v en decision frame w orks. He is co- author of the 2025 MDPI Drones paper “Multi-Agent Actor–Critic Frame w orks for U A V Sw arm Net- w orks” (Q1, IF 4.4), In this research, His contrib ution in this research is conceptualizat ion, method- ology , softw are, v alidation, formal analysis, in v estig ation, resources, data curation, writi ng – original draft, re vie w and editing, visualization, project administration. He is committed to creating intelligent distrib uted systems and describing theoretical models as concret e, reproducible research outputs. He can be contacted at email: shan26103@gmail.com. Multi-model deep ensemble fr ame work for early dia gnosis of r ar e g enetic ... (Shan Mahmood) Evaluation Warning : The document was created with Spire.PDF for Python.
224 ❒ ISSN: 2502-4752 Sayma Akte r T rina is a B.Sc. student in Computer Science and Engineering at the Amer - ican International Uni v ersity-Bangladesh, with a strong academic focus on AI, ML, and data-dri v en systems. Her research interests include ML, DL, algorithm optimization, and sentiment and emotion analysis using generati v e models. In the current study , she contrib uted to the methodology design, In v estig ation, V isualization, softw are, Data curation, formal analysis, v alidation, and w as acti v ely in v olv ed in bot h the original draft pr eparation and manuscript re vie w and editing. She is dedicated to b uilding interpretable and scalable intelligent systems that inte grate theoretical adv ancements with real-w orld applications. She has been r ecognized with the Dean’ s A w ard for outstanding academic performance. She can be contacted at email: trinasayma5191@gmail.com. Ar pita Saha Suk anna is an Ame rican International Uni v ersity-Bangladesh B.Sc. Com- puter Science and Engineering student. In this research, her contrib ution is methodology , softw are, data curation, and writing original draft. Her areas of research are ML, DL, and NLP . She is ac- ti v ely e xpanding her impact in AI through joint research ef forts and ongoing publications. She can be contacted at email: arpita.sukanna85@gmail.com. Sabrina Zaman Esha is an under graduate student in the Department of Computer Science and E ngineering at Ame rican International Uni v ersity Bangladesh. Her research inter ests include NLP and AI-dri v en c ybersecurity . She has been in v olv ed in research projects focusing on enhanc- ing email security using DL and NLP-based models. Her contrib ution is visualisation, data curation and writing original draft. She is passionate about using intelligent systems to address real-w orld problems in digital communication. Her academic contrib utions continue to gro w through collabo- rati v e research and publications in the eld of AI and c ybersecurity . She can be contacted at email: needbasic51@gmail.com. Md. Agdam Amin Adib is currently studying Computer Science and Engineering (CSE) at American International Uni v ersity-Bangladesh. He is an under graduate student. His research in- terest is in applying ML and DL to healthcare to solv e real-w orld medical challenges by utilizing Python, P andas, scikit-learn, and basic genomic data processing tools. His research e xpertise in- cludes ML, DL, and biomedical data analysis with a focus on healthcare solutions and genomic data applications. His contrib ution to this research is softw are and writing the original dr aft. He aims to contrib ute practical, research-based solutions that can mak e a meaningful impact in healthcare to contrib ute to the betterment of humanity . He can be contacted at email: agdam.adib@gmail.com. Md. Sanim Ahmed is a B.Sc. student in Computer Science and Engineering a t the Amer - ican International Uni v ersity-Bangladesh with research interests in ML, DL. He has contrib uted to this research in the follo wing capacities as a co-author of the 2025 MDPI Drones paper “Multi-Agent Actor–Critic Frame w orks for U A V Sw arm Netw orks” (Q1, IF 4.4). In the current study , contri- b utions inc lude writing the original draft. His w ork focuses on creating interpretable and scalable intelligent systems that connect theory to real-w orld implementation. He can be contacted at email: ahmedsanim1234@gmail.com. Amirul Islam recei v ed his P .hD. de gree in Computing and Electronic Systems from the Uni v ersity of Esse x, UK, in 2022. He completed his M.Sc. from K ookmin Uni v ersity , South K orea. He currently serv es as an Assistant Professor in the Department of Electrical and Elec- tronic Engineering at the BRA C Uni v ersity , Bangladesh. Prior to this, he held the position of a Post-Doctoral Researcher at the V isual AI Laboratory , Oxford Brook es Uni v ersity , UK. His research interests include ML for communication, optical camera communication, deep reinforcement learn- ing, automoti v e v ehicular communications, and optimiza tion strate gies. He can be reached at email: amirul.islam@bracu.ac.bd. Indonesian J Elec Eng and Comp Sci, V ol. 42, No. 1, April 2026: 215–224 Evaluation Warning : The document was created with Spire.PDF for Python.