Our Products

DETECTOS

Imagine a situation where hundreds of medical scans, tests and images are screened by AI, and only the ones with abnormalities are presented to the experts. Other than the increased speed in diagnosis, medical experts will be able to take on more work and focus specialized abilities on high value, decision-making work.

CONTINUUM Enterprise

Hospital performance, both operational and clinical, can be dramatically increased when data are used to detect areas of performance bottlenecks. Simulation of new management strategies will result in precise decision making and policy setting. Guesswork in hospital management will be a thing of the past.

CONTINUUM Population

Managing for optimal health in situations of limited resources is what Public Health is mostly about. Being able to quantify and predict the impact of public health policies where large sums of funds are being allocated would serve as a powerful tool for public health administrators and healthcare payers of any size of population.

CONTINUUM Clinical

Multiple factors impact the outcome of certain medical interventions ranging from behavioral change to medication to surgery, including individual patient differences. Machine learning on the effects of different treatment options matching key patient characteristics and attributes could lead to the desired state of Personalized Medicine.

Our Approach

Listen

We listen to your problems, your concerns, how you operate, and what matters most to you and your organization.

Look

We take a good look at your data. Quality and quantity of your data, including how it is stored, are important in telling us how to address your issues.

Link

We, and preferably with you, think of desired outcomes, measurements and applicable solutions. Formulating the question is a key part of success.

Learn

We train the machine. The machine will absorb all the knowledge from your organization from the data fed to it. It will see patterns. It becomes intelligent.

Leverage

We, after satisfactory testing, help you deploy it in real-world situation, measure if it reaches the desired objective, tweak it for increased performance, and more tweaking until you are happy.

Our Projects

Detectos: Automated Abnormality Screening for Chest X-rays

Detectos: Automated Abnormality Screening for Chest X-rays
The metric that matters

Screening Performance at Radiologist-Level Sensitivity

AUC is a useful metric when comparing between machine learning models but it does not address how effective a model is for screening.

We use precentage of image screend with radiologist-level sensitivity of 92.6% (%saved) as our performance metric.

Accuracy does not matters
Accuracy 2/10 = 20%
Sensitivity 2/3 = 66%
%saved 5/10 = 50%
Positive Predictive Value 2/5 = 40%
Optimizations Architecture Public Data Local Data %saved AUC
CheXNet/Pert densenet121 112k NA 23% 0.71
CheXNet/Pert densenet121 112k 47k 47% 0.86
BridgeAsia densenet121 112k 47k 59% 0.90

Predicting Readmission in Diabetic Patients

Predicting Readmission in Diabetic Patients
How do we train our model

Predict Next Patient Visit within the 5-10 Days Average Error

Health Check-up Data
Physical examination, lab results and other check-up data about a patient.
Patient Visit Data
All unplanned visit with ICD codes by a patient duing the last year.
Disease Grouping
Our medical doctor group ICD codes into five major disease groups.
Predict Next Visit
We predict the likelihood of visiting a hospital within the next year.

Deep Neural Networks with Mixed Inputs

We use embedding layers to extract features from catagorical data, combine them with numerical data, and model them through deep neural networks.

Target Disease Groups Mean Absolute Error (Days)
Obesity 5.68
Hypertension 10.85
Diabetes 11.10

Frequently Asked Questions

With modern machine learning, especially deep learning, techniques, we there are three main types of problems: detection, prediction and automation. Detection involves identifying specific group characteristics of your samples, be it medical images, lab results or patient profiles. Prediction allows us to know in advance the likely clinical and financial outcomes of each operation. And from these two, we can automate decisions according for optimal results.
It depends. There are several rule of thumps based on extrapolation from past research such as at least 20,000 images per label for classification and at least 100,000 samples of tabular data for deep learning approach to outperform traditional machine learning. However, it is best to bring your sample dataset and discuss with us.
Each solution offers its advantages and disadvantages. It all depends on the company's culture and the data. Cloud offers a perfect option for companies that need deployment simplicity, secure processing, and easily scalable platform. It is perfect for those who need the full feature of A.I. with little deployment cost and manpower. On-Premise, however, is perfect for those with sensitive content and a very restrictive legislative environment. It offers the ultimate solution in security, privacy, and compliance.
There are many concerns when handling patient data for machine learning in healthcare, for instance, data cleansing, data storage, data privacy and security etc. However, in our view, the biggest concern is the data privacy and security. Therefore, we use HIPAA standard when handling the patient data with all of our clients.
The most difficult part of machine learning in healthcare is preparing the data to be both for the model and for HIPAA compliance. If you can, with our support, clear these requirements, we will take care of setting up your machine learning models and train your personnel on how to maintain and tune them on your own.
Before you measure performance, you need to understand your priority. For instance, if you want to use the model for screening, you would want to achieve the highest sensitivity possible even though you might have a relatively high false positive rate. On the other hand, if you are concerned about false positives, you might want to use a more balanced metric like the F1 score.
In order to customize and maintain the model, you would need a team with backend engineers who can transform your datasets into HIPAA-compliant format our models ingest as well as distribute the model outputs to end users. As for performance tuning, domain owners such as clinicians can tune the models directly from our user interface without any programming expertise.

Contact

Leave us a message