AI-based Studies in Epidemiology & RCS.pptx

AI-based study in epidemiology

1. AI-enabled (Deep learning-based) Studies in Epidemiology and, Reproductive and Child Health Department of Public Health & Informatics  Bangabandhu Sheikh Mujib Medical University, Dhaka Presented By: Dr. Khandaker Reajul Islam MS in Orthodontics from BSMMU, Dhaka MSc in Physiology from UKM, Malaysia BDS from Dhaka Dental College Resident at Department of Orthodontics, BSMMU, Dhaka.

2. Machine learning (ML) is a type of artificial intelligence (AI), that allows, software applications, to become more accurate at predicting outcomes, without being, explicitly programmed to do so. Supervised machine learning models are trained with labeled data sets, which allow the models to learn and grow more accurate over time. For example, an algorithm would be trained with pictures of cats and other things, all labeled by humans, and the machine would learn ways to identify pictures of cats on its own. Machine learning (ML)

3. Difference between classic Machine Learning and deep learning Techniques Ref.- https://www.dynam.ai/wp-content/uploads/2020/06/Screen-Shot-2020-06-26-at-11.04.18-AM-e1593376707667.png

4. Classical Machine Learning vs Deep Learning Algorithms A wide range of classical machine learning models was employed. Such as: logistic regression (LR), support vector machine (SVM), XGBoost Random Forest (RF) K-nearest Neighbors (KNN) Ada Boost, etc. A small amount of Deep learning techniques were also used, which includes: Convolutional Neural Network (CNN) Yolo variants HRNet variants

5. How object detection model works? Ref.- https://www.analyticsvidhya.com/blog/2018/12/practical-guide-object-detection-yolo-framewor-python/18-AM-e1593376707667.png YOLO- You only look once Deep learning object detection Model- by Joseph Redmon

6. How anatomical landmark can be detected? Ref.- https://medium.com/codex/facial-landmark-detection-algorithms-5b2d2a12adaf Facial landmark detection Deep learning based anatomical landmark detection model

7. Deep Learning-based Application Development for Automated Cephalometric Analysis

8. Study plan 8 Ref: https://www.nature.com/articles/srep33581/figures/1

9. Study Design with DL Models 9 Cropped according to the center location and resized to 512x512 Predicted heatmap after 2 nd stage Input Image Cropped according the center Location and resized to 256x256 Predicted heatmap Patch Images Extraction 1 st Stage SelfCephaloNet Model SelfCephaloNet Model 2 nd Stage

10. Study Design with DL Models 10

11. CephaloNet (BSMMU Server) Doctor’s Portal Cephalometric Image Acquisition System X-ray Operator X-ray Images X-ray Images Results Results X-ray Images Web-Interface 11

12. Our Active Web Page Active link: https://orthoai.netlify.app/ 12

13. 13 Demonstration of Ortho AI

14. Generated Result by Our Apps 14

15. Early Prediction of Sepsis Using Machine Learning Techniques 15

16. Infection Inflammatory Response Dysregulated Immune Response Endothelial Dysfunction Coagulation Abnormality Organ Dysfunction Septic Shock Pathophysiology of Sepsis 16

17. Clinical Criteria of Identifying Patients with Sepsis and Septic Shock https://www.researchgate.net/profile/Vinayak-Patki-2/publication/327425121/figure/fig2/AS:667182325313538@1536080070128/Operationalization-of-Clinical-Criteria-Identifying-Patients-with-Sepsis-and-Septic.png 17

18. I have access to publicly available data from two U.S. hospital: Beth Israel Deaconess Medical Center and Emory University Hospital. It consists of : Total ICU data of about 40,000 participants includes 40 clinical variables: 8 vital sign, 26 laboratory variables, and 6 demographic variables, those of which at least were recorded once in 2 days of an ICU stay . The time series include multiple observations that were written in hours and minutes after admission to ICU. Vital Signs HR O2Sat Temp SBP MAP DBP Resp EtCO2 Database Description Lab Values BaseExcess Glucose HCO3 Lactate FiO2 Magnesium pH Phosphate PaCO2 Potassium SaO2 Bilirubin_total AST TroponinI BUN Hct Alkalinephos Hgb Calcium PTT Chloride WBC Creatinine Fibrinogen Bilirubin_direct Platelets Demographics Age Gender Unit1 Unit2 HospAdmTime ICULOS 18

19. Positive labels of sepsis were found in about 3000 of the 40,000 records, which is 7.27% of the data. sepsis labels were found in about 2800 rows, which is only 2% of all data. As the ratio of sepsis to non-sepsis labels are very imbalanced, we have to be careful to create strategies to handle these imbalanced data. Features of Data Sets 19

20. Study Design with Classical ML Classical ML Models: MLP Classifier ElasticNet Linear Discriminant Analysis XGB Classifier Random Forest Classifier Logistic Regression ExtraTrees Classifier ADABoost Classifier KNN Classifier and Gradient Boosting Classifier Feature Ranking Techniques: XGBoost Random Forest ExtraTrees FiO2/SaO2 BUN/Creatinine 20

21. Study Design with Classical ML Stacking Machine Learning Technique 21

22. Feature Ranking 22

23. Conclusion Due to extremely high data imbalance, standard classical ML based approach and stacking ML models are performing poor for the early prediction of sepsis. Although negative class is predicting by the model with very high specificity, but the precision, recall, and F1 score are poor. Splitting the negative class into 3, 5 or 10 splits and carrying out 5-fold cross-validation has significantly improved the overall performance. SMOTE and COPULA augmentation with split method showed that for 12-hour earlier than clinical onset of sepsis, COPULA augmentation perform much better than SMOTE; however, SMOTE performs well for 24-, 36-, and 48-hour onsets. To evaluate whether this observation holds in real-world scenario or not, we need to conduct an external validation in the future using a prospective study. 23

24. AI-enabled (Deep learning-based) Application Development for Diabetic Foot Ulcer Management .

25. DFUs account for 84% of lower limb amputations worldwide, The Wagner scoring method is commonly used for DFU severity assessment, though differentiating between certain stages (e.g., Grade 2 and Grade 3) is challenging with visual images alone. Diabetic foot ulcers Ref: https://biocomposites.com/eu/wp-content/uploads/sites/2/2023/02/Case-Study-2-Block-1.png

26. The Wagner Scoring Method The commonly used classification system for diabetic foot ulcers. Based on depth, infection, and the presence of gangrene.   S ix grades to classify the ulcer:  Grade 0 : No ulcer, but a high-risk foot (presence of deformities, calluses, or erythema). Grade 1 : Superficial ulcer involving the skin only. Grade 2 : Deeper ulcer involving the subcutaneous tissue, which may expose bone, tendon, ligament, or joint capsule. Grade 3 : Deep ulcer with abscess, osteomyelitis (bone infection), or joint sepsis. Grade 4 : Localized gangrene (partial foot gangrene). Grade 5 : Extensive gangrene involving the whole foot.

27. Primary Objective: To develop a deep learning (DL) model for the accurate classification of the severity of diabetic foot ulcers (DFUs) using a combination of visual and thermal imaging along with clinical history data. Secondary Objectives: To compare the accuracy of the AI-based model with conventional methods, such as visual inspection and thermal asymmetry analysis, in classifying DFU severity according to the Wagner scale. To assess the reliability of the AI model in predicting DFU progression across different patient demographics, such as age, sex, and underlying health conditions (sub-analysis). Objectives of My Research

28. Thermal and Visual images of a Male patient with type 2 DM Ref: https://link.springer.com/article/10.1007/s13340-017-0315-1

29. 1.Background and Literature Review C onduct a comprehensive literature search using academic databases such as PubMed and IEEE Xplore to identify studies related to AI applications in DFU detection, diagnosis, and severity stratification. A nalyze the challenges in the existing literature These findings will form the basis of our study rationale and will be summarized into a cohesive narrative that supports the proposed research . Study Design

30. 2. Creation of a Benchmark Dataset: Creating the first-ever benchmark dataset for diabetic foot ulcers including visual and thermal foot images from patients, along with their clinical history. The collected dataset will be classified into six grades based on the Wagner scale. Grade 2 Grade 3 Grade 4 T he anonymized dataset will be used for AI model training, validation, and future diabetic foot ulcer research initiatives. Study Design

31. Study Design (Self-Organized Operational Neural Networks) 3. AI Model Development for DFU Classification

32. BD Diabetic Foot Care App 4. AI-enabled Mobile Application for DFU Monitoring

33. AI-enabled (Deep learning-based) Application Development for Early Prediction of Pre-eclampsia .

34. What is Pre-eclampsia A serious medical condition that can occur about midway through pregnancy (after 20 weeks). People with preeclampsia experience- High blood pressure (more than 140/90) Protein in their urine (more than or equal to 300 mg/24 hours ) Swelling Headaches Blurred vision

35. Height Weight Gestation Period Age Manually Entry Data Data set creation

36. Blood Pressure Protein Level Heart Rate Respiration Rate Clinical Data Data set creation

37. Data set creation

38. PPG signals PPG Respiration Rate BP Data set creation

39. Data set creation

40. API (Application Programming Interface) is a set of rules and protocols for building and interacting with software applications. It allows different software systems to communicate with each other. In the context of the document, the API is used to upload medical data from smart biomedical devices to a central database and to create notifications for healthcare workers and patients. ​ Study Design

41. Conclusion This research is to facilitate more smart biomedical devices in the healthcare sector Save more lives by an early prediction generated by machine learning. Machine learning approaches can extract different parameters of diagnosis from one data. Appropriate and immediate notification can be generated from these data using state-of-the-art machine learning approaches. Using this preeclampsia management system the pregnant women of rural areas will be more beneficial.

42. Validation of a Deep Learning-Based Chest X-ray Application (TB- CXRNet ) for Tuberculosis Detection in Bangladesh

43. It’s a form of tuberculosis resistant to anti-TB drugs due to incomplete or improper treatment, as well as misuse of TB drugs, which allows the bacteria to adapt and become resistant. There are primarily two types of drug-resistant TB MDR-TB (Multidrug-Resistant TB) : This type of TB is resistant to at least the two most effective first-line TB drugs: isoniazid and rifampicin . XDR-TB (Extensively Drug-Resistant TB) : This is a more severe form of MDR-TB. It is resistant not only to isoniazid and rifampicin but also to any fluoroquinolone and at least one of the second-line injectable drugs (like amikacin, kanamycin, or capreomycin). XDR-TB is even more difficult to treat than MDR-TB due to its resistance to a broader range of drugs. Both types pose significant treatment challenges, requiring longer, more complex, and potentially more toxic treatment regimens. Drug-resistant TB:

44. Study Design Phase- B Phase- A 93.32% 87.48% 79.59%

45. https://www.kaggle.com/datasets/raddar/drug-resistant-tuberculosis-xrays?select=tbportals_metadata.csv Study plan

46. References [1]  Hossain MB, Khan MN, Oldroyd JC, Rana J, Magliago DJ, Chowdhury EK, Karim MN, Islam RM. Prevalence of, and risk factors for, diabetes and prediabetes in Bangladesh: Evidence from the national survey using a multilevel Poisson regression model with a robust variance. PLOS Global Public Health. 2022 Jun 1;2(6):e0000461. [2] Goyal M, Reeves ND, Rajbhandari S, Yap MH. Robust methods for real-time diabetic foot ulcer detection and localization on mobile devices. IEEE journal of biomedical and health informatics. 2018 Sep 6;23(4):1730-41. [3] Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. Communications of the ACM. 2017 May 24;60(6):84-90.. [4] Šavc, Martin, Gašper Sedej , and Božidar Potočnik . "Cephalometric Landmark Detection in Lateral Skull X-ray Images by Using Improved SpatialConfiguration -Net." Applied Sciences 12, no. 9 (2022): 4644. [5] Kwon, Hyuk Jin , Hyung Il Koo, Jaewoo Park, and Nam Ik Cho. "Multistage Probabilistic Approach for the Localization of Cephalometric Landmarks." IEEE Access 9 (2021): 21306-21314. [6] Lindner, Claudia, Ching -Wei Wang, Cheng-Ta Huang, Chung- Hsing Li, Sheng-Wei Chang, and Tim F. Cootes . "Fully automatic system for accurate localisation and analysis of cephalometric landmarks in lateral cephalograms ." Scientific reports 6, no. 1 (2016): 1-10. [7] https://figshare.com/articles/dataset/dental-cepha-dataset_zip/13265471/1 accessed on 12 July 2022. [8] https://webceph.com/en/ accessed on 12 July 2022. [9] Wang, Ching -Wei, Cheng-Ta Huang, Jia -Hong Lee, Chung- Hsing Li, Sheng-Wei Chang, Ming- Jhih Siao , Tat-Ming Lai et al. "A benchmark for comparison of dental radiography analysis algorithms." Medical image analysis 31 (2016): 63-76. [10] Lindner, Claudia, and Tim F. Cootes . "Fully automatic cephalometric evaluation using random forest regression-voting." In IEEE International Symposium on Biomedical Imaging (ISBI) 2015–Grand Challenges in Dental X-ray Image Analysis–Automated Detection and Analysis for Diagnosis in Cephalometric X-ray Image. 2015.

47. Thank You