Applied Medical-Informatics: Transforming Healthcare Research with AI and Signal Processing

A Collection of Research in Applied Medical Informatics:
1. Methods for Continuous Blood Pressure Estimation
2. Immune Mechanisms in Liver Transplantation: Linking Etiology, Ischemia-Reperfusion Injury, and Graft Survival
3. Knowledge-Enhanced Multimodal Learning for Improved Disease Diagnosis Generation
4. Integrating Time-Series and Text for Advanced Patient Monitoring with MedTsLLM
5. Enabling Personalized Healthcare with Improved Multi-View Data Imputation
6. Achieving High Dice Scores in Liver Pathology Detection with RMA-Mamba
7. SegMamba: Speed and Accuracy in 3D Medical Segmentation
8. Improving CDSS (Clinical Decision Support System) Accuracy for Diabetic Retinopathy with Spectral Features

1. Practical Applications in Medical
Informatics Research
By Sione Palu






Contents
1. Methods for Continuous Blood Pressure Estimation
2. Immune Mechanisms in Liver Transplantation: Linking
Etiology, Ischemia-Reperfusion Injury, and Graft Survival
3. Knowledge-Enhanced Multimodal Learning for Improved
Disease Diagnosis Generation
4. Integrating Time-Series and Text for Advanced Patient
Monitoring with MedTsLLM
5. Enabling Personalized Healthcare with Improved Multi-
View Data Imputation
6. Achieving High Dice Scores in Liver Pathology Detection
with RMA-Mamba
7. SegMamba: Speed and Accuracy in 3D Medical
Segmentation
8. Improving CDSS (Clinical Decision Support System)
Accuracy for Diabetic Retinopathy with Spectral Features
9. References

2. 1. Methods for Continuous Blood Pressure Estimation
• Problem: Traditional methods for continuous blood pressure monitoring, such
as those based on PTT or PPG waveform parameters, suffer from poor
generalization and require frequent individual calibration due to variations in
signal shape caused by age and disease.
• Solution: A novel, non-invasive hybrid deep learning model uses a combination
of Ensemble Empirical Mode Decomposition (EEMD) and a Temporal
Convolutional Network (TCN) to predict blood pressure from PPG signals.
• Advantages: This approach automates the extraction of complex signal features,
eliminating the need for manual feature selection. It addresses the limitations of
traditional machine learning methods and offers improved accuracy over other
deep learning models like CNN- and RNN-based hybrids, as shown by its lower
prediction error.
• Significance: This model provides a more robust, generalized, and accurate
solution for continuous, non-invasive blood pressure monitoring.
• Read more in the summary provided in Ref [1] at the end: Methods for
Continuous Blood Pressure Estimation Using Temporal Convolutional Neural
Networks and Ensemble Empirical Mode Decomposition

2. Immune Mechanisms in Liver Transplantation: Linking
Etiology, Ischemia-Reperfusion Injury and Graft Survival
• Problem: The connection between ischemia-reperfusion injury (LIRI) during liver
transplantation and subsequent immune responses is poorly understood.
Existing methods lack the ability to comprehensively analyze complex,
longitudinal datasets of immune markers and liver function tests, hindering the
identification of distinct immune profiles and their link to clinical outcomes.
• Solution: A novel tensor-based factorization approach was developed to
synthesize complex longitudinal data from liver transplantation, effectively
identifying distinct immune profiles from cytokine levels and liver function tests.
This method can model the temporal dynamics of the immune response over the
course of the transplant.
• Key Finding & Advantage: The refined model identified two immune profiles that
are predictive of five-year allograft survival. These profiles are influenced by
clinical factors like donor/recipient age and underlying disease. This provides a
clear link between immune responses, LIRI, disease etiology, and long-term graft
survival, which was not previously established.

3. • Clinical Impact: This approach offers actionable clinical insights for
personalizing post-transplant care. By identifying the specific immune
imbalances (e.g., in CD4+ T-cell polarization) that increase the risk of graft
failure, it helps pinpoint potential therapeutic targets tailored to the patient and
donor characteristics to improve transplant outcomes.
• Read more in the summary provided in Ref [2]: Tensor-Based Integration of
Time-Series Measurements Reveals Relationships Between Underlying
Disease, Early Injury, and CD4+ Polarization in Liver Transplantation

3. Knowledge-Enhanced Multimodal Learning for Improved
Disease Diagnosis Generation
• Problem: Current disease prediction models in medical informatics primarily
handle single data types (modalities), ignoring the wealth of information
available in diverse electronic health records (EHRs). Existing multimodal
approaches use separate encoders, which fail to fully integrate information into a
unified feature space and lack clinically relevant insights, limiting their predictive
accuracy.
• Solution: The EHR-KnowGen model is a novel approach that integrates multiple
EHR modalities (structured and unstructured data) into a unified feature space.
It achieves this by combining a multimodal learning model with a large language
model (LLM) and leveraging soft prompts for effective data fusion.
• Advantages: Unlike previous methods, EHR-KnowGen enhances the extraction
and fusion of multimodal data by incorporating external domain knowledge at
various levels of detail. This leads to improved prediction performance and
provides explainable evidence for the generated diagnoses, enhancing clinical
understanding and trust.
• Result: The model demonstrates superior performance over existing methods on
real-world EHR datasets, proving its effectiveness in creating more accurate and
interpretable disease diagnosis predictions.
• Read more in the summary provided in Ref [3]: EHR-KnowGen - Knowledge-
enhanced multimodal learning for disease diagnosis generation

4. Integrating Time-Series and Text for Advanced Patient
Monitoring with MedTsLLM
• Problem: Analyzing complex, multi-modal patient data—especially diverse
physiological time series and text—for clinical decision support is highly

4. challenging. Traditional methods require meticulous feature engineering and
struggle to effectively integrate and interpret information from different data
types, limiting their ability to provide actionable insights.
• Solution: The MedTsLLM framework is a novel multimodal large language model
(LLM) designed to directly integrate physiological time series data with rich
textual context. It uses a reprogramming layer to align raw time series with the
LLM's embedding space, eliminating the need for extensive manual feature
engineering.
• Advantages: This approach enables deeper analysis of patient data and
performs three critical clinical tasks: semantic segmentation, boundary
detection, and anomaly detection. By tailoring prompts with patient-specific
information, it provides a powerful tool for generating actionable insights.
• Results: MedTsLLM significantly outperforms state-of-the-art baselines across
various medical domains, demonstrating superior accuracy in tasks involving
electrocardiograms (ECG) and respiratory waveforms.
• Read more in the summary provided in Ref [4]: MedTsLLM - Leveraging LLMs for
Multimodal Medical Time Series Analysis

5. Enabling Personalized Healthcare with Improved Multi-View
Data Imputation
• Problem: A major challenge in analyzing longitudinal omics data for
personalized healthcare is incomplete data views, as current methods are not
well-suited to accurately impute this missing information. This hinders future
predictions and the exploration of crucial temporal dynamics.
• Solution: LEOPARD is a novel method designed to solve the problem of missing
data in multi-timepoint omics datasets. It operates by separating the data into
content and temporal components, using the temporal information to
intelligently and accurately fill in the missing content.
• Advantages: LEOPARD significantly outperforms traditional imputation methods
such as missForest, PMM, GLMM, and cGAN. The data it imputes shows stronger
agreement with observed data, leading to more reliable analyses of age-
associated and disease-related biological markers.
• Impact: By providing a robust solution for handling missing views, LEOPARD
enables a more comprehensive analysis of longitudinal omics data, offering a
significant contribution to personalized healthcare and a deeper understanding
of an individual’s changing physiology over time.

5. • Read more in the summary provided in Ref [5]: LEOPARD - missing view
completion for multi-timepoint omics data via representation
disentanglement and temporal knowledge transfer

6. Achieving High Dice Scores in Liver Pathology Detection
with RMA-Mamba
• Problem: Traditional Transformer architectures are computationally inefficient
for long sequences (e.g., high-resolution medical images) due to their quadratic
complexity, limiting their application in tasks like automated liver segmentation.
This process is currently manual, time-consuming, and prone to error.
• Solution: The RMA-Mamba architecture leverages the efficiency of State Space
Models (SSMs), which have a linear complexity, making them suitable for
processing long sequences. It builds on this foundation by adding a novel
Reverse Mamba Attention (RMA) module to effectively capture both local details
and global context.
• Advantages: This dual approach allows RMA-Mamba to process high-resolution
medical images with computational efficiency, while also accurately modeling
complex features and long-range dependencies. It automates the segmentation
process, providing a consistent and reliable alternative to manual methods.
• Performance: RMA-Mamba achieves state-of-the-art performance in
pathological liver segmentation from both MRI and CT scans. It demonstrated
high accuracy on the new cirrhotic liver dataset (CirrMRI600+), and a very high
Dice score of 92.9% on the cancerous liver segmentation dataset (LiTS).
• Read more in the summary provided in Ref [6]: A Reverse Mamba Attention
Network for Pathological Liver Segmentation

7. SegMamba: Speed and Accuracy in 3D Medical
Segmentation
• Problem: The Transformer architecture, while powerful for capturing long-range
dependencies, is computationally expensive and memory-intensive, especially
when processing high-dimensional data like 3D medical images. This limits its
practical application for tasks such as whole-volume segmentation.
• Solution: SegMamba is a novel 3D medical image segmentation model that
leverages the efficiency of the Mamba State Space Model (SSM) architecture.
This approach enables it to effectively capture long-range dependencies across

6. an entire image volume with significantly improved computational speed and
memory efficiency.
• Advantages: Unlike Transformer-based models, SegMamba can process
complex 3D medical images with much greater speed, even at a high resolution
of 64x64x64. This makes it a more scalable and practical solution for clinical
applications that require rapid analysis.
• Performance: Experiments on the BraTS2023 dataset confirm that SegMamba is
highly effective and efficient, demonstrating its superiority over existing methods
for whole-volume segmentation.
• Read more in the summary provided in Ref [7]: SegMamba: Long-range
Sequential Modeling Mamba For 3D Medical Image Segmentation

8. Improving CDSS (Clinical Decision Support System)
Accuracy for Diabetic Retinopathy with Spectral Features
• Problem: Existing automated methods for Diabetic Retinopathy (DR) diagnosis,
primarily based on Convolutional Neural Networks (CNNs), often overlook
spectral features in fundus images. These features contain valuable quantitative
information (e.g., vessel diameter, retinal thickness) crucial for accurate
detection and grading, which is a key limitation of relying solely on spatial
features.
• Solution: A new Clinical Decision Support System (CDSS) model was
developed that leverages both spatial and spectral features from retinal fundus
images. The model uses a Wavelet CNN to capture multi-scale details and
extract a robust feature vector, which is then classified by a Support Vector
Machine (SVM) to determine the severity grade of DR.
• Advantages: The inclusion of Wavelet CNN allows the model to analyze images
at multiple scales, preventing the loss of important fine-grained details. By
incorporating spectral features, the system achieves higher accuracy in
detecting and classifying DR severity levels. This approach improves upon
existing methods by providing a more comprehensive analysis of retinal
characteristics.
• Performance: Experimental results on the EyePACS dataset show that the
proposed Wavelet-CNN-SVM model outperforms other state-of-the-art
techniques across multiple performance metrics, including precision, recall, F1-
score, accuracy, and AUC score. This demonstrates its effectiveness in providing
a more reliable and efficient tool for early DR diagnosis.

7. • Read more in the summary provided in Ref [8]: Classification of Diabetic
Retinopathy Disease Levels by Extracting Spectral Features Using Wavelet
CNN
9. References
1. “Methods for Continuous Blood Pressure Estimation Using Temporal Convolutional
Neural Networks and Ensemble Empirical Mode Decomposition”,
• https://www.linkedin.com/posts/sione-palu-3803b0bb_medicalinformatics-
python-r-activity-7089424541722017792-ajeZ/
2. “Tensor-Based Integration of Time-Series Measurements Reveals Relationships
Between Underlying Disease, Early Injury, and CD4+ Polarization in Liver
Transplantation”,
• https://www.linkedin.com/posts/sione-palu-3803b0bb_python-activity-
7365511263130406912-NhiG/
3. “EHR-KnowGen - Knowledge-enhanced multimodal learning for disease diagnosis
generation”,
• https://www.linkedin.com/posts/sione-palu-3803b0bb_medicalinformatics-
activity-7155057252624691200-BKpw/
4. “MedTsLLM: Leveraging LLMs for Multimodal Medical Time Series Analysis”,
• https://www.linkedin.com/posts/sione-palu-3803b0bb_medicalinformatics-
python-activity-7351725306526814208-S2Ca/
5. “LEOPARD - missing view completion for multi-timepoint omics data via
representation disentanglement and temporal knowledge transfer”,
• https://www.linkedin.com/posts/sione-palu-3803b0bb_medicalinformatics-
python-activity-7355743901204295680-i9_m/
6. “A Reverse Mamba Attention Network for Pathological Liver Segmentation”,
• https://www.linkedin.com/posts/sione-palu-3803b0bb_medicalinformatics-
python-activity-7361170356058906624-rEz-/
7. “SegMamba: Long-range Sequential Modeling Mamba For 3D Medical Image
Segmentation”,
• https://www.linkedin.com/posts/sione-palu-3803b0bb_medicalinformatics-
python-activity-7286915728572665857-qkLO/
8. “Classification of Diabetic Retinopathy Disease Levels by Extracting Spectral
Features Using Wavelet CNN”
• https://www.linkedin.com/posts/sione-palu-3803b0bb_medicalinformatics-
activity-7203951574006591489-KA1p/