We develop AI-enabled applications to improve medical diagnosis and treatment.
We aim to
Our team specializes in data harmonization of medical data and mapping it to the OMOP Common Data Model, enabling standardized analysis and characterization of healthcare data. With our extensive experience in this domain, we have successfully worked with a wide range of medical datasets from diverse sources, including electronic medical records (EMR) and claims data. One of our key strengths lies in the comprehensive mapping of medical codes to OMOP Standard Concepts. We meticulously transform diverse code systems, including standard terminologies, proprietary codes, and local coding systems, into the standardized OMOP vocabulary. Our meticulous approach ensures accurate and consistent representation of medical concepts, enabling seamless integration and interoperability across different data sources. Once the data is mapped to the OMOP Common Data Model, we seamlessly integrate it into your existing analytics toolset. Whether you utilize OHDSI tools or commercial applications, our team ensures smooth incorporation of the standardized data.
Medical image processing refers to a set of procedures performed in order to extract clinically meaningful information from various imaging modalities (MRI, CT, etc.) mainly for diagnosis or prognosis. The modalities are typically in vivo types.
Machine Learning models are used in order to compute relevant features. After being trained with the prior knowledge provided by medical experts, Machine Learning guides image registration, fusion, segmentation, and other computations towards accurate descriptions of the initial data and extraction of reliable diagnostic cues.
MRI
| Magnetic Resonance (MR) images play a major role in current medical and research settings. The acquisition process of MRI induces inherent noise and artifacts, hence, filtering methods are usually required to improve the image quality. We offer different tools for performing MRI denoising either as a preprocessing step to a further analysis or in order to provide images of higher quality to physicians. The different methods used can be categorized into “classical” Machine Learning (ML) models using the intrinsic pattern redundancy of the image patterns and those exploiting their sparseness properties and into Deep Neural Networks (DNN) which use supervised learning by training different architectures with pairs of noisy and noise-free data and try to infer the clean image from the noisy input. Different types of artifacts exist in different MRI modalities and some of them are very difficult to be detected and removed, due to the fact that they are very rare or quite similar to regular components (e.g. the eye artifact in fMRI, or the thermal drift). For artifact removal different Blind Source Separation (BSS) methods can also be applied in order to successfully remove any potential artifact. |
Denoising and artifact removal
|
Image registration is the process of transforming different images into one common coordinate system so that common features overlap and differences, if any, between the different images can be emphasized and become visible, even to the naked eye. There are a host of clinical applications requiring image registration. For example, one would like to compare two MRI scans of a patient, taken at different time instances in order to quantify any possible difference between the two, e.g., the growth of a tumor during the intervening time period or the different activated areas in fMRI. We offer different tools for affine intra- and inter- model registration. Although many registration methods have been proposed and can be used, it is still a challenging task in some scenarios, such as images with large anatomical variations, multimodal registration, etc. in such cases DNNs are employed to provide the optimal registration of the medical images. |
Image registration
|
The expanding number and quality of the images debilitates the radiologist’s abilities for translating them. Computer vision-based applications of biomedical imaging are gaining more importance as they provide recognition information to the radiologist for better treatment-related problems. The formation of abnormal groups of cells, inside the brain or near it, leads to the initialization of a brain tumor. The abnormal cells abrupt the processing of the brain and affect the health of a patient. Various features such as image texture, eigenvalues, smoothness of the shape or local histograms can be used by ML methods, such as Support Vector Machines (SVMs) and Random Forest (RF) for the lesion segmentation and delineation. Furthermore, deep-learning-based techniques are becoming popular in lesion segmentation studies, as their performance can be superior if they are trained with a proper amount of data. |
Quantification of brain tumours
| Epilepsy is one of the most common neurological diseases and more than 50 million people worldwide suffer from epileptic seizures. Ιn some of the epileptic patients epilepsy is caused by a structural cause, such as a scar or a lesion on the brain. That’s the reason MRI, and other structural imaging modalities are used to look for potential structural causes and ML tools can help into this goal. Various features such as image texture, eigenvalues, smoothness of the shape or local histograms can be used by ML methods, such as Support Vector Machines (SVMs) and Random Forest (RF) for the lesion segmentation and delineation. Furthermore, deep-learning-based techniques are becoming popular in lesion segmentation studies, as their performance can be superior if they are trained with a proper amount of data. |
Quantification of epileptic lesions
fMRI
| The localization of the activated brain areas is a challenging Blind Source Separation (BSS) problem, in which the sources consist of a combination of spatial maps (areas activated) and time-courses (timings of activation). Our brain is not activated only from the task information we process, but also from external and (sometimes) even unconscious stimuli (e.g. heart beating, breathing, thinking of personal and family issues not connected to the task, etc.), hence the separation of the activation directly connected to the hand in task is a difficult and complex problem. With the use of Function M R I(fMRI) and different BSS tools like tensor methods, or Dictionary Learning (DL) or Independent Component Analysis (ICA) we can perform different types of analysis in order to identify the activated areas in single or multi-subject studies. |
Mapping of the brain activation
|
The study of connectivity of the brain is an important factor in order to better understand how the human brain works and to delineate areas that need to stay untouched during a brain surgery in order to prevent severe counter-effects. Structural (or anatomical) connectivity refers to the existence and structural integrity of tracts connecting different brain areas (i.e. white matter tracts connecting cortical areas/nuclei), while MRI and diffusion MRI are being used in order to study and visualize it. Functional and effective connectivity are neuroimaging terms. Functional connectivity refers to the statistical dependence – correlation of the signal from different brain areas (usually with the use of resting state fMRI). Effective connectivity is a bridge between the two different connectivity measures and brings in the element of causality (i.e. a signal, activation in one area directly causes a change or signal, activation or depression, in another area). We can offer different connectivity analysis and visualizations based on the task at hand. |
Functional and effective connectivity
Diffusion MRI
| The study of connectivity of the brain is an important factor both to better understand how the human brain works and to delineate areas that need to stay untouched during a brain surgery in order to prevent severe counter-effects. Structural (or anatomical) connectivity refers to the existence and structural integrity of tracts connecting different brain areas (i.e. white matter tracts connecting cortical areas/nuclei), while MRI and diffusion MRI are being used in order to study and visualize it. Normalized probability distribution and Orientation Distribution Function (ODF) provide information about the distribution of water diffusion in the brain as a function of direction. Probabilistic or Deterministic Direction Getters can be used in order to estimate the tracts of neurons, along which, the water molecules propagate in the brain. |
Effective connectivity and tractography
Ultrasound
|
AINIGMA Technologies have created an automated deep learning workflow to classify, segment, and annotate two-dimensional (2D) videos and Doppler modalities in echocardiograms. This workflow aims to replace the time-consuming and error-prone manual interpretation of echocardiograms, which is currently used to assess cardiac systolic and diastolic function for heart failure diagnosis and management. |
Automated Cardiac Ultrasound Assistance Tool
CT
|
We offer different tools for performing CT denoising either as a preprocessing step to a further analysis, or in order to provide images of higher quality to physicians. The different methods used can be categorized into “classical” ML models using the intrinsic pattern redundancy of the image patterns and those exploiting their sparseness properties and into DNNs which use supervised learning by training different architectures with pairs of noisy and noise-free data and try to infer the clean image from the noisy input. Different types of artifacts exist in CT and some of them are very difficult to be detected and removed due to the fact that they are very rare. For artifact removal different Blind Source Separation (BSS) methods can also be applied in order to successfully remove any potential artifact. |
Denoising, artifact removal
| Stroke is the leading cause of serious long-term disability and the fifth leading cause of death worldwide. Non contrast Computed Tomography (CT) of the brain remains the mainstay of medical imaging in the setting of an acute stroke since it is the gold standard for detecting hemorrhage. Machine learning tools along CT images can be used both for quantifying the extent of the stroke (with similar methods used in brain tumor quantification in MRI), while also for predicting which patients will develop Symptomatic Intracranial Haemorrhage (SICH). |
Quantification and prediction of stroke
Biomedical signal processing involves the treatment and analysis of bio-signal measurements. By using machine learning tools, the signals can be analyzed in order to help physicians with greater insights to make better decisions in clinical assessments.
The combination of novel machine learning models with signal processing methods has assisted in revealing information which entirely altered the diagnosis and prognosis of several diseases.
The application-oriented and data-driven bio-signal analysis and processing applications, not only benefit from the machine learning algorithms, but also encourage the development of novel (bio)medical treatments.
EEG
|
Worldwide, more than 70 million people suffer from epilepsy, a neurological condition characterized by seizures that affect their quality of life. Automated seizure warning and detections systems can be used to improve the patients’ quality of life. During a seizure different neurons of the brain are activated rapidly and without any external stimuli. This activation of the neurons and the release of ions between the neural synapses can be measured via electroencephalography (EEG). Different features can be extracted from the EEG signals which will help us cluster the time points where a seizure takes place. The computation of those handcrafted features is the first step in “Classical” ML approaches. The disadvantage of such methods is the fact that prior knowledge of the physiology of the problem is needed in order to extract those features. On the other hand, DNNs can be employed. The use of NNs facilitates the automatic identification of relevant features, but a big amount of data is necessary in order to train such models. |
Seizure detection, seizure prediction
| Sleep disorder is a symptom of many neurological diseases that may significantly affect the quality of daily life. Furthermore, sleep in infants plays an important role in brain maturation, and this could be of great interest in the case of preterm infants. Hence, it can be understood, that studying the sleep stages of subjects is of critical importance for many biomedical applications. Visual labelling of the sleep stages is time consuming, and difficult so automated analysis of EEG to identify sleep stages is of great interest to clinicians. Different approaches can be explored, the state-of-the-art machine learning tools with handcrafted features are commonly used for the sleep stage classification but the rapid fluctuations between sleep stages often result in blurry feature extraction, hence, approaches based on NNs are proposed. |
Sleep stages classification
| Brain-Computer Interface (BCI) is an interface which establishes an one-to-one communication between the human brain and the computer device to control. EEG signals other than being used for medical reasons can also be used for BCI. Currently, the BCI systems provide two types of signals, raw signals and logic state signals. The latter signals are used to turn on/off the devices. One of the bigger problems in BCI is that the brain signals are weak and fluctuate a lot, hence, the design of a classifier is a hard problem. Human experience and prior knowledge can help on certain aspects but fall insufficient in more general circumstances. An automatic feature extraction method is highly desirable and that’s the reason why neural networks are highly used in BCI applications. |
Brain-Computer Interface (BCI)
ECG
| Electrocardiogram (ECG) is an important diagnostic tool for the assessment of cardiac arrhythmias in clinical routine. ECG is the electrical representation of the contractile activity of the heart, and can be recorded fairly easy. The rhythm of the heart in terms of beats per minute can be easily calculated by counting the R peaks of the ECG wave during one minute of recording. Thus, heart rhythm disorders or alterations in the ECG waveform are evidence of underlying cardiovascular problems, such as arrhythmias. The detection of arrhythmias can also be used in the detection of epilepsy, anxiety or other pathological conditions. based on the accuracy, computing power or the speed needed by the application, different approaches can be explored for the detection of arrhythmia. |
Arrhythmia detection
|
Worldwide, more than 70 million people suffer from epilepsy, a neurological condition characterized by seizures that affect their quality of life. Despite the different treatments currently available, still 35% of the patients continue to experience seizures for the rest of their lives Automated seizure warning and detection systems can be used to improve the patients’ quality of life. Based on the type of the seizure that the patient exhibits, multimodal algorithms can be employed by using data from different sources. In patients with focal seizures a combination of ECG and EEG is optimal. It has been exhibited that for some subjects the use of ECG alone is adequate, and since mobile ECG is easier to be measured, such an option is preferred. We offer different types of models and algorithms that can be used for accurate seizure detection and prediction. Classical ML approaches can be used with handcrafted features obtained from expert knowledge, while also DNNs can be employed. |
Seizure detection, stages of anxiety detection
EMG
| Worldwide, more than 70 million people suffer from epilepsy, a neurological condition characterized by seizures that affect their quality of life. Despite the different treatments currently available, still 35% of the patients continue to experience seizures for the rest of their lives Automated seizure warning and detection systems can be used to improve the patients’ quality of life. Based on the type of the seizure that the patient exhibits, multimodal algorithms can be employed by using data from different sources can be collected. In patients with tonic clonic seizures a combination of EMG and EEG is optimal. We offer different types of models and algorithms that can be used for accurate seizure detection and prediction. Classical ML approaches can be used with handcrafted features obtained from expert knowledge, while also DNNs can be employed. |
Seizure detection, examination of electrical activity of muscles, determination of muscle fatigue
Infectious diseases are caused by microorganisms belonging to the class of bacteria, viruses, fungi, or parasites. Despite the advances in medicine, infectious diseases are a leading cause of death worldwide, especially in low-income countries.
With the advent of Machine Learning tools, we can provide more accurate and in-time prediction of epidemics, understand the specific features of each pathogen, and identify potential targets for drug development. Machine Learning tools and methods are able to identify patterns in data, provided that adequate data are available.
Patient medical data are increasingly being stored as electronic health records (EHRs) at healthcare institutions worldwide. However, EHRs can be inconsistent and noisy, may contain many missing values, and frequently include unstructured text fields. Nevertheless, the very fact that these data are electronically available in large volumes provides the potential for applying Machine Learning, including in the field of infectious disease management.
 | Machine learning algorithms are designed to discover patterns and associations in data; they learn from that data and encode that learning into a model to accurately predict a data attribute of importance for new data. With high-quality data, subtle nuances and correlations can be accurately captured and high-fidelity predictive systems can be built. Real World Data (RWD) denotes patient level data which is collected from different sources as part of routine healthcare practice, hence RWD is seen as a potentially rich source to enhance the understanding of a patient’s disease trajectory and to generate insights as to how clinical care affects patient outcome under real world conditions, with the aid of machine learning tools. |
Use of real world data for medication optimization
|
Sepsis is a major global health problem which is estimated to affect about 49 million people annually and cause about 11 million related deaths. Apart from the serious death toll, the burden of sepsis is further increased by its various social and economic implications, since sepsis is a major cause of increased length of in-hospital stay, readmission and disability. Different algorithms can be trained by processing real-world clinical and laboratory data and their variations with the aim to develop a sensitive tool that will predict the risk of sepsis in admitted patients and detect sepsis incidents early during the disease, hence raising the awareness of healthcare providers, to intervene as soon as possible. |
Sepsis prediction
| An infectious disease occurs when a person is infected by a pathogen that eventually causes disease. Infectious Diseases are, in most cases, communicable, therefore they can be transmitted in various ways. Much of the destructive potential of infectious diseases stems from the fact that they often strike unexpectedly, leaving little time for preparation. The best countermeasure is therefore an early warning to give affected regions or communities more time to prepare for the impact. We think that such a system for classifying infectious disease risk would help to guide the development of infectious disease intelligence and to identify best courses of action. |
Prediction of the spread, outcome of an infectious disease
Technologies we use:
