Blockchain and explainable-AI integrated system for Polycystic Ovary Syndrome (PCOS) detection

PCOS identification based on machine learning and deep learning

A recent study by Subha et al. (2024) concentrated on PCOS detection with the use of ML algorithms. For feature selection, Swarm Intelligence (SI) was used, whereas for disease prediction, random forest (RF) and XGBoost classifiers were used. Both Flashing Firefly (FF) and Particle Swarm Optimization (PSO) feature selection algorithms yielded 92% accuracy. In the research work by Danaei Mehr & Polat (2022), different feature selection methodologies were used on the same dataset. The model’s efficacy was analyzed with various ensemble classifiers for both feature reduced dataset and feature not-reduced dataset. This study concluded that Ensemble random forest classifier performed well along with embedded feature selection in comparison with other classifiers. A wide collection of ML algorithms had been applied on the tabular non-invasive parameters and their performances have been analyzed in Tiwari et al. (2022). Random forest was chosen as the best model with 93.25% accuracy.

Ravishankar et al. (2023) adopted a fuzzy-based convolutional neural network (CNN) method for determining the presence of ovarian cysts. This technique yielded good results with 98.37% accuracy rate. An efficient PCOS detection method was proposed by a research study found that Extra trees classifier, in combination with Genetic algorithm for feature selection, gave better results for PCOD disease detection in comparison with other classifiers (Munjal & Khandia, 2020). Nandipati, Chew & Wah (2020) performed out an analysis with RapidMiner and Python Scikit package tools using several classifications and feature selection techniques for the same dataset, they found that RapidMiner tool performed better. An app for PCOS diagnosis was developed by an earlier research study, analyzing several boosting and bagging algorithms (Kavitha et al., 2023). It was observed that the proposed web app facilitated prediction utilizing Grid search cross-validation (CV) optimization with an accuracy of 94%.

Suganya et al. (2023) used the Pelvic Computed Tomography (PCT) dataset to detect ovarian cysts. The features were extracted using GoogleNet and the output was supplied to XGBoost algorithm for further classification. When compared to deep learning networks, this approach proved to be more accurate and efficient. There were a few additional studies that focused on using deep learning’s transfer learning models like VGG16, MobileNet, Inception v3, U-Net and Fusion techniques to analyze ultrasonic imaging data for PCOS (Vikas, Radhika & Vineesha, 2021; Lv et al., 2022; Abouhawwash et al., 2023; Alamoudi et al., 2023).

Blockchain and machine learning integration

An earlier study (Hu & Kar, 2024) focused on the development of a secure health monitoring system using blockchain technology and distributed ML. Hyperledger Fabric was utilized for user management and grouping, and ModelChain was used for model training. Sharma & Rohilla (2023) developed a framework with blockchain and ML for drug discovery chain management. The system’s efficiency was improved with an increase in endorsers and block sizes, and comparative study further validated the scalability of the architecture. The system’s performance was successfully assessed using the throughput and model correctness.

A system for detecting the severity of cardiovascular disorders by combining the naïve Bayes algorithm and Hyperledger Fabric was proposed by Sasikumar & Karthikeyan (2023). Using this system, physicians could easily determine the patient’s illness severity level and quickly identify high-risk patients. The designed system outperformed the existing system based on the comparison of central processing unit (CPU) and memory utilization. Heidari et al. (2023) addressed the need for timely lung cancer detection by using blockchain-based federated learning (FL) to identify the malignant nodules. The obtained outcomes revealed that the CapsNet classification was effective in identifying lung cancer patients, with 99.69% accuracy and minimum classification errors. Pandey et al. (2022) proposed a system for waste material identification and recommendations for reducing the quantum of disposed waste with its reuse. The system utilized deep learning’s ResNet50 model for exact object identification. Upon successful object identification, the system leveraged a Web Scraper to get applicable recycling ideas from the internet. This approach encouraged environmental sustainability and facilitated environmental pollution reduction.

A reliable system for veterinary data storage and analytics was proposed by Iqbal et al. (2021) it could make several predictive analyses to improve veterinary services, using regression and deep learning models. Implemented with Hyperledger Fabric in two variations, a single Kubernetes cluster and in a cross-cluster environment, the proposed system employed advanced algorithms to accurately forecast appointment scheduling frequencies, leading to optimal resource allocation and increased operational efficiency in veterinary healthcare settings. A natural language processing (NLP) based model integrated with blockchain to convert the doctor’s prescription into text was presented by Bharimalla et al. (2021). This system used Hyperledger Fabric to build the EHR system, integrated with national identity documents like Aadhaar or PAN (Permanent Account Number) to easily identify the patient’s medical records.

The reviewed studies on PCOS detection and Blockchain-ML integration present a variety of approaches, each with its strengths and limitations. Subha et al. (2024) achieved 92.64% accuracy using Swarm Intelligence for feature selection and random forest (RF) and XGBoost classifiers but lacked blockchain integration for secure data management. Danaei Mehr & Polat (2022) have proposed Ensemble random forest classifiers that integrate feature selection from within; however, this has improved performance but doesn’t cover scalability and blockchain security. Ravishankar et al. (2023) have obtained a high accuracy of 98.37% with the fuzzy-based CNN for the detection of ovarian cysts. They propose a strong deep learning-based approach, yet they don’t involve blockchain. Suganya et al. (2023) applied GoogleNet for feature extraction and XGBoost for classification, which outperformed other deep learning models but was not equipped with data security measures. In blockchain-ML integration, Sharma & Rohilla (2023) showed how Hyperledger Fabric and ModelChain improved scalability and throughput in secure health monitoring systems, though this approach was not designed for PCOS detection. Overall, although these studies show promising results in the detection of diseases and security on blockchain, they are often ignored when both technologies are used in combination for safe, transparent, and efficient diagnosis of PCOS. This proposed work seeks to bridge this gap.

Based on the extensive literature review, it is observed from Table 1 that there is a deficit of research studies that have evaluated both blockchain and ML metrics with their integration. Most of the works, achieved a high accuracy but were not generalizable to diverse populations, thus raising concerns about their applicability in clinical settings. Moreover, the issues of data quality, model complexity, and lack of external validation limit the robustness of findings. Research which integrates explainable AI is mostly seen in early phases only but proper approaches in transparency modeling of the same is mostly seen missing in previous literatures and thus, also diagnostic mainly as opposed to a combination solution. It is perceived that there is a distinct lack of a transparent, safe, blockchain-based mechanism for classifying PCOS disease. This proposed work aims to address the identified knowledge gap by utilizing blockchain and AI powered PCOS diagnosis for precise PCOS detection.

System architecture

The architecture of the proposed EAIBS-PCOS system is given in Fig. 1. The blockchain module is designed with Hyperledger Fabric, which is a private consortium network. The health data is collected from the patients through multiple sources, as shown in Fig. 1 and the collected data is supplied as input into the Hyperledger Fabric network. Figure 2 illustrates the stakeholders contributing to the EAIBS blockchain module. These stakeholders form the base organizations of the fabric network and their active peers are involved in endorsing, chaincode design and deployment. Every entity is set with certain privileges and only authorized participants can interact and access data in this permissioned blockchain network. The clients register themselves with the EAIBS-PCOS system through the gateway to gain access to the information.

The client applications that submit transactions make their requests through the fabric gateway and the respective smart contracts are fetched. A transaction submission can be made to start a prediction analysis, obtain specific data, submit new data or request modifications to existing data. For PCOS prediction, the chaincode designed will gather an adequate amount of the patient’s data and pass it to the ML module. After data processing, model training and predictive analytics, the results are sent back to the blockchain module. The results are recorded in the ledger, so that it remains uncorrupted and are then passed back to the requested client applications. The results are not available to the public, thereby ensuring complete confidentiality of patient health records. Only the stakeholders permitted by the patients can access their personal and predictive data.

The key stakeholders involved in the EAIBS-PCOS system, and their work limitations are given below,

1. Patient-People who have a health history and request medication. They can control and update their own profiles, change their passwords, view health information, grant, withdraw or deny access to other stakeholders.

2. Physicians-Medical professionals working with hospitals, they diagnose the ailments of patients, devise treatment plans and update patient records. They can manage their own records and must request permission from patients to access their medical records.

3. Insurers-Insurance companies who make insurance updates, payments and maintain case history are permitted to view the relevant part of the physician assistants (PA’s) records.

4. Health center-Hospitals maintain clinical data of patients and profiles of all the stakeholders.

5. Researchers-Laboratories focus on research studies for developing more effective treatment approaches, maintain records of controlled experimental trials evaluate test outcomes and encourage the conduct of clinical research to evaluate the effect of new drugs and potential treatment options.

6. Pharmacy-Medical retailers who retain patient medication information and add specifics about their medications.

7. Public health agency-Regulatory authority which manages vaccination history, gathers demographic data about the patients, examines patient’s public records and devises community health development plans.

Hyperledger fabric components

Figure 3 depicts the network architecture of the EAIBS-PCOD system’s Hyperledger Fabric design. It’s important components are explained in brief, as follows.

PCOS ML predictive analytics

Blockchain and ML unified prediction system

In the PCOS predictive analytics system, after the authorization and authentication, the client application makes the request for transaction submission through the gateway and the transaction proposal gets submitted to the blockchain module. It then fetches the corresponding chaincode and passes the data to the ML module. The query transactions details are passed to the ML module, and it acts as the backend. The ML algorithms used for the proposed design are the support vector machine, K-nearest neighbours, random forests, XG Boost, and naive Bayes.

Support vector machine

A supervised learning algorithm, support vector machine (SVM) locates the optimal line or hyperplane in an N-dimensional space to optimize the distance between each class (Kecman, 2005). In order to maximize the margin among the nearest data points in each class—known as support vectors—SVM first finds the hyperplane that best divides the data into discrete classes. By using kernel functions, which translate the input data into higher-dimensional spaces where it becomes linearly separable, the approach is able to handle both linear and non-linear data (Suthaharan, 2016; Pisner & Schnyer, 2020).

K-nearest neighbours

It is an instance-based learning technique. Based on the majority class of its K nearest neighbors in the feature space, the K-NN algorithm classifies a data point. Because it functions on the premise that identical data points are nearby, it is very simple when dealing with issues involving patterns that are spatial or distance-related (Mucherino, Papajorgji & Pardalos, 2009; Zhang, 2016).

Random forests

It is a popular ensemble learning technique. During training, it builds a number of decision trees, each of which is trained using a different subset of features and data at random, via combining the findings from each individual tree, the final prediction is produced. Because specific decision trees are prone to overfitting and variance, this technique increases the accuracy and resilience of the model (Pal, 2005; Rodriguez-Galiano et al., 2012; Biau & Scornet, 2016).

XGBoost

A refined version of gradient boosting methods, XGBoost (Extreme Gradient Boosting) is intended for use in both regression and classification applications. It produces an ensemble of decision trees, where each new tree seeks to repair the flaws of the preceding ones by optimizing a differentiable loss function (Chen, 2015). Because of a number of enhancements, including regularization to lessen overfitting, parallelization of the tree-building process, and internal management of missing data, XGBoost is renowned for its efficiency, scalability, and high performance (Chen & Guestrin, 2016; Dhaliwal, Al Nahid & Abbas, 2018).

Naive Bayes

It is based on Bayes’ Theorem. Because of the class label, it is predicated on the idea that a dataset’s features are conditionally independent of one another. The data point is assigned to the group with highest likelihood after the algorithm determines the probability of each data point belonging to that class and it works well for many real-world scenarios (Murphy, 2006; Rish, 2001; Berrar, 2025).

To obtain a better performance, several models are blended based on the analysis of the results. The best models are ensembled and finalized for PCOS prediction and analysis. This ensemble model is integrated with Explainable AI (XAI) algorithm to ensure transparency in prediction. Figure 4 explains all the workflow behind the proposed EAIBS-PCOS system. It is clear that the ML module processes the data and with the data of the patient, analysis is done using the finalized ensembled learning algorithm. It is then integrated with XAI algorithms, and the results are then passed back to the blockchain module. Now, the clients can access the prediction results through the gateway.

Performance evaluation setup

Table 3 gives the components set up for the proposed system’s implementation. In addition, an interactive computing platform, Jupyter notebook is utilized to ensure training, testing and validation of various machine learning strategies for the PCOS detection system. Table 4 lists the parameters set to complete the evaluation of the proposed design.

Dataset

An open-source dataset is taken from the Kaggle database (Kottarathil, 2020) and the owner of this dataset, Kottarathil P had collected physical and clinical data of 541 women patients from multiple hospitals in Kerala. A wide range of 41 attributes such as body mass index (BMI), follicle number, age, weight, height, pulse rate, etc., are in the chosen dataset.

Data preprocessing

As a part of the data preprocessing, inconsistent and null values are eliminated. The missing values are identified and replaced with the median value. A few unwanted attributes such as Serial no. and Patient File no. are removed. Several attributes such as BMI, follicle-stimulating hormone (FSH)/luteinizing hormone (LH), waist/hip ratio are rounded off to two decimal places. Features with similar characteristics are removed using the Kendall correlation approach. Figure 5 presents the results of Kendall’s Method. Feature scaling is done using Min-Max Scaling to regulate the dataset’s range of independent variables. The formula for performing Min-Max Normalization is given by,

(5) $$x^{\mathrm{\prime }}={\displaystyle \frac{x-min\left(x\right)}{max\left(x\right)-min\left(x\right)}}$$

The Synthetic Minority Oversampling Techniques (SMOTE) is used to ensure the balancing of data in both training and testing phases. A 70:30 split between training and testing sets is made from the balanced dataset. Data analysis is done for better understanding of datasets.

Evaluation metrics

The performance evaluation parameters of EAIBS-PCOD (Ferri, Hernández-Orallo & Modroiu, 2009; Hossin & Sulaiman, 2015; Tharwat, 2018; Hyperledger Blockchain Performance Metrics White Paper, 2018; Pajooh et al., 2022; Malik et al., 2023) are given below. Table 5 gives the details of the confusion matrix. The description about the parameters that are utilized in the evaluation metrics are illustrated in the Table 6.

Read latency (s): It is the time interval between submission of request for reading a transaction and receiving the reply. Equation (6) gives its formula,

(6) $$LR=T_{RR}+T_{RR}-T_S.$$

Read throughput (rps): It is the calculation of how many read operations are finished in a certain amount of time. It is shown in Eq. (7).

(7) $$T{H}_R={\displaystyle \frac{R_T}{T_T}}$$

Transaction latency (seconds s): It is the interval that occurs between submitting a request for a transaction and the time it takes for the result to be distributed throughout the network. It is calculated by using Eq. (8).

(8) $$LT=T_C-T_S$$

Transaction throughput (tps): It is the calculation of the total number of valid transactions completed within a specific time frame. Equation (9) gives its formula.

(9) $$T{H}_T={\displaystyle \frac{T{C}_T}{T_T}.}$$

Resource consumption: It is the measure of resources consumed such as maximum memory, average memory and CPU power, by the components and processes of the Hyperledger network.

Accuracy (Acc): It is the percentage of a model’s correct predictions, or the number of samples properly identified divided by the total number of samples. Equation (10) is used to calculate the Acc.

(10) $$A_{cc}={\displaystyle \frac{t_p+t_n}{t_p+f_p+t_n+f_n}.}$$

Precision (P): It is the measure of positive predictions that are successfully forecasted from the total predictions. It is calculated by using Eq. (11).

(11) $$P={\displaystyle \frac{t_p}{t_p+f_p}.}$$

Recall (R): It is the percentage of positive patterns that are accurately identified. It is shown in Eq. (12).

(12) $$R={\displaystyle \frac{t_p}{t_p+t_n}.}$$

F1-score (F1): It is the harmonic average of recall and precision levels. The formula for calculating the F1-score is given in Eq. (13).

(13) $$F1={\displaystyle \frac{2.P.R}{P+R}.}$$

By employing a comprehensive set of performance measures—accuracy, precision, recall, F1-score, AUC, throughput, latency, and transaction execution time—this study ensures a thorough evaluation of the proposed model’s effectiveness. These metrics not only provide insights into predictive performance but also address the practical considerations of system efficiency and user experience, making them essential for assessing the overall impact of the integrated Blockchain and ML solution in healthcare.

Hyperledger fabric’s performance assessment

The fixed-rate controller of the Hyperledger Caliper framework is used to evaluate the reliability of the planned blockchain network. The fixed-rate controller is simple to configure, generates a stable workload and enables scalability testing. Under this, three parameters (tps (transactions per second), Workers, transaction number (TxNum)) are varied, and their corresponding latency and throughputs are recorded for analysis. Initially, three experimental cases are analyzed, as given below, and the findings are discussed. For each reading in the experimental cases, the experiment is repeated ten times and it is averaged.

Case 1-Varying the tps

Tps indicates the speed at which the cumulative transactions of all workers are sent to the system under test (SUT). From Fig. 6A, it is noted that the read latency remains the same (0.02 s) and transaction latency increases with an increase in tps, but the variation has no major impact. As seen in Fig. 6B, with an increase in tps of 25 for each round, it is observed that there is a clear and steady linear increase in read throughput which is proportional to the considered tps. The transaction throughput increases linearly up to tps of 75 and after that it is in the same throughput range of 80 to 81 tps. As depicted in Fig. 6C, with an increase in tps, time taken for execution decreases for both read and create operations.

Increasing the tps enables faster completion of transactions and increases the throughput with no major impact on latencies, as seen in Fig. 6. From this analysis, it is observed that choosing maximum tps of 125 will result in optimized network, with acceptable latency levels.

Case 2-Varying the workers

Workers replicate the actions of actual clients or users participating in the blockchain network, and are responsible for creating and reporting the transactions during the benchmarking. Figure 7 gives an analysis of graphs depicting the variation of workers from 25 to 100. Figure 7A clearly depicts that the read latency is either 0.02 or 0.03 s and transaction increases linearly from 0.27 to 0.61 s. In case of throughputs, Fig. 7B shows a slight increasing pattern for both the cases. Figure 7C reveals that the execution time of transactions decreases with an increase in workers.

A large number of workers will lead to parallel processing and maximizing the throughput, as seen in Fig. 7. To optimize the throughput while ensuring acceptable latency and execution time with the consideration of the present study’s findings, a manageable workforce size of 100 could be chosen.

Case 3-Varying the TxNum

TxNum represents the number of total transactions to be executed at each round during the benchmarking. With the experimentation of varying TxNum from 500 to 2,500, from Fig. 8A it is found that both transaction and read latency remain constant with the values of 0.02 and 0.26 s, respectively. As shown in Fig. 8B, both transaction and read latency remain almost the same, within the range of 25 to 25.2 tps. Figure 8C shows that the execution time increases linearly with increasing TxNum.

Figure 8 shows that higher TxNum decreases the throughput and increases the execution time. Therefore, lower values of TxNum will help in improvement of the system’s performance. Hence, an optimum value of 500 TxNum is chosen for the designed system.

From the analysis made above, the optimized network parameters are identified. This optimized network is utilized for further integration with the PCOS machine learning module.

PCOS detection using XAI

Client enrolment and ledger query

Admin enrollment and client registration are the essential processes to query and access the service of the proposed system. Figure 9A portrays its results. After the successful enrolment, the client submits a query for making the transaction request proposal. Figure 9B gives the results of the query submitted by the client that displays the profile of a particular patient based on a unique patient ID.

Through the call.js file shown in Fig. 10A, the prediction request is passed to the ML module, processed over there and the results are sent back to the client. Figure 10B shows the results of the prediction query made by the client through the gateway. Here, the prediction results are requested for the patient with ID 4. The patient’s prediction result is 0, which implies that the patient is not diagnosed with PCOS. Simultaneously, the blockchain ledger will also be updated accordingly.

Performance comparison of ML models

The prediction results from ML models are based on pre-trained models which are chosen and deployed in the Jupyter notebook. Figure 11 gives the receiver operating characteristic (ROC) curves of the chosen ML algorithms. The ROC curve is a graphical illustration of the efficacy of a binary categorization strategy. The ROC curve is used to measure the model’s ability to discriminate between positive and negative categories. Usually, the curve begins at (0, 0) and finishes at (1, 1) (Zweig & Campbell, 1993; Bewick, Cheek & Ball, 2004). Figure 12 gives the confusion matrix of the chosen ML models.

Based on the results presented in Fig. 12, RF and XGBoost algorithms are identified to perform best and they are ensembled further to enhance the performance. Figure 13 gives the comparison chart of various ML algorithms and Ensembled algorithms. This figure clearly shows that the Ensemble model is the best for the chosen problem statement with 98.04 accuracy, 99.5 precision, 96.5 recall and 97.9 F1-score with 10-fold cross validation. Further the standard deviation of 0.01 is acquired for measured accuracy, indicating the robustness of the model across all the runs.

The computational efficacy is of the proposed approach is assessed with its training and testing time. The model’s appropriateness for real-time applications is demonstrated by its 1.38-s training time and 0.01-s testing time. Hence due to its relatively shorted computational time, the model is more relevant for time-sensitive health application like PCOS diagnosis with guaranteed security and scalability.

XAI prediction results

Figure 14 gives the force plot of SHAP algorithm results that are applied on two test dataset instances and the contribution of different attributes on the model’s output is clearly depicted. Features in the red sections of the text plot enhance the model’s output when included, whereas sections that are blue lower the model’s output when included. The values at the top, 0.51 in Fig. 14A and 0.72 in Fig. 14B represent the model’s prediction probability.

A list of explanations that show how each feature influences the prediction of the considered instance is given in Fig. 15 with the use of LIME algorithm and the attributes highlighted in orange contribute to the score of probability 1. Those attributes highlighted in blue contribute to the score of probability 0. This makes the forecast simpler to understand locally and makes it possible to recognize the feature changes that will have the most effects on the prediction. From Figs. 15A & 15B, it noted that the predicted probability is 1 with the probability score of 0.75. From Figs. 15C & 15D, it is evident that the predicted probability is 0 with probability score of 0.84.

Comparison with state-of-art works

The performance comparison of the suggested EAI-PCOS system with previous research in terms of precision, F1-score and accuracy is shown in Fig. 16. With the use of an ensembled RF classifier, Tiwari et al. (2022) were able to achieve an accuracy of 93.25% with 98.28 precision and 95.4 F1-score. Subha et al. (2024) achieved a maximum accuracy of 92.64% with PSO feature selection and RF classifier. They achieved 91.84% precision and 91.37% F1-score. Rahman et al. (2024) used the mutual information model, with AdaBoost (AB) and RF classifiers, achieving an accuracy of 94%, F1-score of 89% and precision of 96.67%. Silambarasan, Nirmala & Mishra (2024) worked on both image and text dataset for PCOS diagnosis, in which 97.5% accuracy, 85% F1-score and 90% precision was obtained. Hassan & Mirza (2020) achieved 96% accuracy, precision and F1-score with the usage of RF classifier.

In comparison with all the other classifiers, the proposed EAI-PCOS system achieves the highest accuracy of 98.04%, highest precision of 99.5% and highest f1-score of 97.9% with 10-fold cross validation. Even though there are many past studies on PCOS diagnosis, this proposed work is the first to integrate blockchain and XAI leading to enhanced security and transparency in decisions.