Periodic Reporting for period 1 - AI-CARE (Thermography AI - cornerstone in future preventive healthCARE)
Berichtszeitraum: 2023-07-01 bis 2024-02-29
It is estimated that more than 230M people worldwide suffer from Peripheral Arterial Disease (PAD) and each year more than 22M of them develop Critical Limb Ischemia - a life-threatening condition with 50% combined incidence of amputation and/or death in a 3-year period. Timely PAD diagnosis is the most important factor to avoid complications by treatment procedures (revascularisation) and proper disease management (daily walks, change of dietary habits, etc.). Unfortunately, half of the PAD-suffering population is asymptomatic and therefore lately diagnosed. Moreover, patients with Diabetes Type 2, who ironically have the highest risk of developing complications as a result of the disease, are often undertreated as current non-invasive diagnostic methods such as Ankle-Brachial Index (ABI) give high percentage false negatives due to a condition called Medial Arterial Calcification. Studies reveal that diabetics bear 7x higher risk of amputation than non-diabetic PAD patients.
The high incidence of PAD-related complications mostly comes from the fact that there are no mass-screening programmes and patients visit a cardiovascular specialist for disease confirmation too late. Pain in the limbs, which is often the first symptom of PAD, leads to very slow triage as patients unnecessarily visit neurologists and orthopedists in more than 30% of the PAD cases. This not only delays proper treatment and increases chances for complications but also negatively impacts healthcare budgets. Again, diebitic patients suffer most from the lack of population-wide screening solutions because of their higher tolerance to pain.
Current PAD diagnostic methods have significant drawbacks. The ABI (comparison between blood pressure in upper and lower limbs) is the most commonly used approach for initial examination, however it is highly unreliable, especially for patients with diabetes. Doppler Ultrasound (DUS) is another non-invasive method for blood flow observation. It has much higher specificity rate than ABI but it is applied very local (per artery) and considered error prone unless performed by a rigorously trained specialist. X-ray Angiography is often used for final examination prior to intervention. It is highly invasive as the patient is injected with a dye (contrast agent) and radiated for about 3 hours. All methods require hospital visits and very often more than one is applied, leading for instance to over-care by 30% of the X-ray Angiographies considered unnecessary.
We decided to apply our long-lasting experience in AI, machine learning, and image recognition to build a diagnostic solution as simple and patient-friendly as the specialist, or the GP, or the Nurse Practitioner, or even the patient's partner, taking a thermal image of him. Using a mobile thermal camera we can capture the heat of the body, detect local blood flow anomalies and notify the healthcare professional or the person involved, if needed.
That way we can enable widely accessible & precise (even for patients with diabetes) rapid non-invasive vascular diagnostics.
The high incidence of PAD-related complications mostly comes from the fact that there are no mass-screening programmes and patients visit a cardiovascular specialist for disease confirmation too late. Pain in the limbs, which is often the first symptom of PAD, leads to very slow triage as patients unnecessarily visit neurologists and orthopedists in more than 30% of the PAD cases. This not only delays proper treatment and increases chances for complications but also negatively impacts healthcare budgets. Again, diebitic patients suffer most from the lack of population-wide screening solutions because of their higher tolerance to pain.
Current PAD diagnostic methods have significant drawbacks. The ABI (comparison between blood pressure in upper and lower limbs) is the most commonly used approach for initial examination, however it is highly unreliable, especially for patients with diabetes. Doppler Ultrasound (DUS) is another non-invasive method for blood flow observation. It has much higher specificity rate than ABI but it is applied very local (per artery) and considered error prone unless performed by a rigorously trained specialist. X-ray Angiography is often used for final examination prior to intervention. It is highly invasive as the patient is injected with a dye (contrast agent) and radiated for about 3 hours. All methods require hospital visits and very often more than one is applied, leading for instance to over-care by 30% of the X-ray Angiographies considered unnecessary.
We decided to apply our long-lasting experience in AI, machine learning, and image recognition to build a diagnostic solution as simple and patient-friendly as the specialist, or the GP, or the Nurse Practitioner, or even the patient's partner, taking a thermal image of him. Using a mobile thermal camera we can capture the heat of the body, detect local blood flow anomalies and notify the healthcare professional or the person involved, if needed.
That way we can enable widely accessible & precise (even for patients with diabetes) rapid non-invasive vascular diagnostics.
We started creating a unique thermal imaging dataset in partnership with multiple hospitals. Providing them mobile thermal cameras and phones and access to our data collection platform, we managed to collect more than 9,500 thermal images of 1,300+ patients.
Together with the medical team of prof. Petrov we explored the application to vascular disease in detail and witnessed that Kelvin Health can be as telling, even positionally, about a severe blood flow problem as the gold standard of X-ray angiography.
And even in some cases when the Doppler ultrasound may show everything is in norm, Kelvin Health was able to detect anomalies that were later confirmed with angiography.
We further developed HIPAA & GDPR compliant software system including an iOS Application, Data collection & annotation platform, and achieved 86% precise AI for Angiosome Segmentation, 95% accurate AI for Quality Control of the input data.
Our sensitivity for PAD detection based on training with data collected via the platform and a selected subset of the patient data is 100% and specificity is 96.2%. In terms of accuracy and f1-score (harmonic mean of sensitivity and precision), the current model achieves 97% and 91% respectively.
We witnessed these extremely high sensitivity and specificity rates when we compare the results of the thermal analysis with those of X-ray angiography and Doppler ultrasound, and we are currently expanding our hospital partnerships to prepare for a multi-site clinical trial.
The leading cardiovascular specialist in the hospitals we are currently working with are extremely excited by the immediate benefits of applying Kelvin Health even during its current research phase.
The major milestones already accomplished:
Part I - before Women TechEU:
- Q3 -2021 Vascular (PAD/CLI) research started in collaboration with Acibadem City Clinic Cardiology and other EU hospitals
- Q2-2022 Advanced thermal data collection & annotation platform operational
- Q3-2022 Carotid arteries hypo-perfusion research started
- Q1-2023 ML-automated quality control of the thermal data operational
Part II - during Women TechEU:
- Q2-Q3-2023 ML-automated leg and head arterial zones segmentation operational
- Q3-2023 Carotid arteries hypo-perfusion research first output
- Q4-2023 Reached 1,000+ patients and 7,900+ thermal images in our thermal dataset
- Q4-2023 Started the CE mark regulatory process
- Q4-2023 Added US specialists as research partners
- Q4-2023 Became member of AHA's CHTI
Together with the medical team of prof. Petrov we explored the application to vascular disease in detail and witnessed that Kelvin Health can be as telling, even positionally, about a severe blood flow problem as the gold standard of X-ray angiography.
And even in some cases when the Doppler ultrasound may show everything is in norm, Kelvin Health was able to detect anomalies that were later confirmed with angiography.
We further developed HIPAA & GDPR compliant software system including an iOS Application, Data collection & annotation platform, and achieved 86% precise AI for Angiosome Segmentation, 95% accurate AI for Quality Control of the input data.
Our sensitivity for PAD detection based on training with data collected via the platform and a selected subset of the patient data is 100% and specificity is 96.2%. In terms of accuracy and f1-score (harmonic mean of sensitivity and precision), the current model achieves 97% and 91% respectively.
We witnessed these extremely high sensitivity and specificity rates when we compare the results of the thermal analysis with those of X-ray angiography and Doppler ultrasound, and we are currently expanding our hospital partnerships to prepare for a multi-site clinical trial.
The leading cardiovascular specialist in the hospitals we are currently working with are extremely excited by the immediate benefits of applying Kelvin Health even during its current research phase.
The major milestones already accomplished:
Part I - before Women TechEU:
- Q3 -2021 Vascular (PAD/CLI) research started in collaboration with Acibadem City Clinic Cardiology and other EU hospitals
- Q2-2022 Advanced thermal data collection & annotation platform operational
- Q3-2022 Carotid arteries hypo-perfusion research started
- Q1-2023 ML-automated quality control of the thermal data operational
Part II - during Women TechEU:
- Q2-Q3-2023 ML-automated leg and head arterial zones segmentation operational
- Q3-2023 Carotid arteries hypo-perfusion research first output
- Q4-2023 Reached 1,000+ patients and 7,900+ thermal images in our thermal dataset
- Q4-2023 Started the CE mark regulatory process
- Q4-2023 Added US specialists as research partners
- Q4-2023 Became member of AHA's CHTI
We currently train the algorithms via a combination of a deep learning algorithm that segments the arterial zones (angiosomes), an algorithm for extracting thermal statistics, and binary classification algorithm based on gradient boosting and decision trees for classifying the statistics in two categories - positive (stenosis or thrombosis has been detected within the given arterial zone) or negative (the arterial zone is healthy).
The medical-grade dataset used for developing the preliminary algorithms consists of data from 42 patients, collected and medically annotated under an observational clinical trial. From these 42 patients, 313 different angiosome statistical data points have been extracted which are used for training the binary classification algorithm.
Total patients: 42
Women: 21%
Men: 79%
Age (min / avg / max): 49 / 67 / 84
Patients’ ethnicity: Caucasian
Patients’ nationality: Bulgaria
Diabetics: 54%
Hypertension: 100%
Smokers: 76%
Previous intervention: 64%
Fontain classification (min / avg / max): 2 / 2.64 / 4 (coding: I-0; II-а-1; II-b 2; III-3; IV-4)
Before / after intervention angiographies: 100% / 100%
“Before” angiography angiosome annotation distribution:
- Healthy - 65%
- Stenosis - 13%
- Thrombosis - 22%
“After” angiography angiosome annotation distribution:
- Healthy - 91%
- Stenosis - 0.7%
- Thrombosis - 8.3%
Positive examples in the lower parts (below the knee) of the legs have higher prevalence than in the upper (above the knee).
The medical-grade dataset used for developing the preliminary algorithms consists of data from 42 patients, collected and medically annotated under an observational clinical trial. From these 42 patients, 313 different angiosome statistical data points have been extracted which are used for training the binary classification algorithm.
Total patients: 42
Women: 21%
Men: 79%
Age (min / avg / max): 49 / 67 / 84
Patients’ ethnicity: Caucasian
Patients’ nationality: Bulgaria
Diabetics: 54%
Hypertension: 100%
Smokers: 76%
Previous intervention: 64%
Fontain classification (min / avg / max): 2 / 2.64 / 4 (coding: I-0; II-а-1; II-b 2; III-3; IV-4)
Before / after intervention angiographies: 100% / 100%
“Before” angiography angiosome annotation distribution:
- Healthy - 65%
- Stenosis - 13%
- Thrombosis - 22%
“After” angiography angiosome annotation distribution:
- Healthy - 91%
- Stenosis - 0.7%
- Thrombosis - 8.3%
Positive examples in the lower parts (below the knee) of the legs have higher prevalence than in the upper (above the knee).