1. Sector: HealthCare
Team Name: Support Vector Humans
Team Members: Soumyadeep Mukherjee, J H M Apoorva, Pratik Agarwal, Abhinandan
Dubey, Rahul Sharma, Otkrist Gupta, Mrinal Mohit, Sai Sri Sathya, Pratik Shah
Challenge: Accessible Healthcare to Billions by providing rapid screening layer between
patient and doctor.
Root-Causes:
Lack of Medical Expertise, Experience:
Today there is dearth of doctors and specialists in India as pointed out by WHO standards
as 1 doctor for every 1700 patients whereas the minimum ratio is 1:1000. At times of
crisis, the patients are asked to shift to higher level facilities due to lack of expertise and
experience of doctors leading to several casualties.
Low Speed of Diagnosis:
Due to lack of preliminary screening layer for patients, doctors today are not able to
prioritize among critical and not so critical patients. Further they spend majority of the
time looking into diagnosis reports which can be automated than treating the patients.
Lack of Accessibility:
Besides shortage of doctors, their unwillingness to work in the rural areas is another issue.
It creates artificial scarcity in one area and high concentration in another. Hence the
situation further worsens in rural, semi-urban areas.
Lack of Standard Framework:
The knowledge transfer among medical fraternity is mostly verbal and a standard
framework for sharing methodologies, new findings, creating a dataset does not exist. In
today’s data driven world, there is a high need to establish a reference framework, dataset
for doctors to collaborate, obtain trends and make future predictions.
Field Trips:
As part of the field trip, we met Dr. Sirish (ENT Specialist), Dr. Roma Bagi (Dentist) and
obtained valuable suggestions for our platform.
Interaction with Dr. Sirish: He used the web interface annotation tool for the ear. He
helped us identify the features that are of interest to the ENT specialists in the dataset we
had, and also shared his experience as a user of the software platform which will help us
update our platform for a perfect user experience. Another important aspect of the
conversation was the challenges that he mentioned are currently faced by doctors in his
field, which can be eliminated to a large extent with the deployment of the automated
platform.
Interaction with Dr. Roma Bagi: We showed her the segmentation tool for automatic
segmentation of plaque affected regions in the mouth and the datasets we have from the
Kumbh Mela which can be used to diagnose caries, gingivitis etc. She made some

2. important remarks about the kind of information that can be extracted from these
datasets and the role of a doctor in validating any abnormalities that the segmentation
tool might detect. She acknowledged the utility of the diagnostic platform as a support
system for the doctors to monitor the progress of a patient.
Solution: To use machine learning over large scale datasets to build automatic screening,
detection and diagnostic tools for doctors, device makers and patients.
Approach:
The accessibility of data pertaining to various organs, vital signs such as blood pressure,
motion sensor data and patient's previous history are to be captured and assessed to
provide accurate diagnosis. Hence Dataset Collection is a crucial aspect. The collected data
will be in the raw form and needs to be accompanied with annotations or diagnostic
reports to make sense of it. In cases where the diagnostic reports are not available, the
data has to be labelled as healthy or unhealthy by providing comments regarding the state
of the condition. Hence a user friendly Data Annotation interface has to be developed to
aid the doctors to label, annotate the data in efficient, hassle-free manner. Once labelled
data is available Machine Learning Models can be trained to provide automatic screening,
detection and diagnostic tools.
Addressing Root-Causes:
Lack of Medical Expertise, Experience:
Automated tools need much lesser or no human intervention once respective machine
learning models are trained successfully. Hence they fill in for lack of medical expertise.
Also with such tools in place, junior level assistants can be efficiently trained by providing
various possible examples in the dataset and their corresponding diagnosis.
Low Speed of Diagnosis:
By providing a rapid screening layer between doctors and patients, preliminary decisions
can be provided by automated tools. The doctor can further assess the condition of the
patient according to the criticality and give more time, attention towards treating and
checking the progress of their health condition. All in all, the doctor's throughput will be
increased with the help of automated tools.
Lack of Accessibility:
Today with emerging mobile health care devices, a preliminary diagnosis of the patients
can be done at rural areas via health camps. With automated tools in place, immediate
diagnosis can be provided and the critical patients can be channelized to specialists for
specific diagnosis and treatment.
Lack of Standard Framework:
Obtaining a diverse dataset to build a generic model in very essential. And a standard

3. framework will aid the process where the data is obtained via secured channels, protocols.
A diverse dataset will act as a reference to the doctors in their practice and they can also
share their methodologies as to how they tackled specific issues.
Proof of Concept:
Data Collection:
Data pertaining to 500 patients was collected during KumbhMela 2015 using mobile health
care devices. AliveCor was used to record ECG, Sopro Care for oral imaging, Cellscope for
Ear imaging, D-EYE for retinal imaging, Kinect Sensors for motion tracking, Vital signs,
Patient's medical history and demographics were collected via questionnaires.
Data Annotation and Labelling:
Doctor’s expertise is required in converting the raw data to a labelled dataset. And towards
this end, user friendly interfaces were built over the course of the camp specific to
obtained datasets.
Ear, Eye Imaging:
Videos pertaining to Ear were focused on ear drum and Eye videos covered the retinal
area. An annotation tool as a web interface was developed to allow the doctor to play,
pause the patient's video and annotate the important areas by labelling them as healthy
and unhealthy with added comments. Additionally replay option, list of annotations were
displayed to review their annotations.

4. Oral Imaging:
Videos of the interior of the mouth were taken in 3 different light conditions which mark
caries (blue light) and gingivitis (yellow light). A segmentation tool was developed which
automatically segments out these regions with the additional feature that gives the doctor
the ability to mark any regions that might have not been detected by the program and
annotate them.

5. ECG Data:
ECG graphs were obtained from AliveCor as pdf files and a processing tool was developed
to obtain numeric data from them.
Motion Tracking:
The body movements such as a subject walking in a straight line, touching his nose or
opening a jar act aid in determining neurological symptoms. A screening tool was
developed where the patient's video was displayed and annotation options were provided
to label the data.

6. Vital Signs:
Vitals contained Galvanic Skin Response of the body for 10 sec and Blood Pressure, Heart
Rate, Oxygen Saturation, Body Temperature, Height and Weight. The GKR was plotted and
other vital signs were tabulated to asses.
Stakeholders:
Device Makers: The medical devices deployed in diagnostic centres, hospitals are mostly
highly expensive due to the high end research, prototyping and testing involved in
developing them. A diverse dataset will benefit them in developing robust devices and
screening, detection, diagnosis tools can be plugged into their devices.
Government: The dataset created via the rapid screening layer, sources from various
doctors will aid in analysis of trends and making predictions. Automatic Diagnosis from
Mobile Health care will enable government health agencies and NGO's to conduct low
cost, effective health camps in rural areas making health accessible to the remote areas as
well.
Doctors: The throughput of the doctors will be significantly increased with the help of rapid
screening layer meanwhile helping them prioritizing the critical vs not so critical patients.
Such a system will help them train the junior level doctors and provide a standard
framework wherein they can contribute to knowledge transfer and also learn new findings
in their respective fields.
Patients: In terms of value being added, the patients will be a major stakeholder in
obtaining available, accessible and affordable health care.

7. Future Work:
During the 7 days of the workshop, several annotation, processing tools were developed
and we received several valuable suggestions, reviews from the specialists regarding the
user interface design. Over the course of coming 6 months, a further iteration in
incorporating the suggestions will be taken up. With annotation tools in place, datasets will
be labelled by expert physicians and machine learning models will be trained for screening,
detecting, automating diagnosis. The end-to-end system would involve several iterations
of the same process to ensure accuracy of the models. We aim to test the developed
models, get reviewed and establish the proof of concept for the future deployment of the
platform at a larger scale.
With accurate models in place, they will be deployed by collaborating with Device
Manufactures, Hospitals and Government. The current dataset involves data from mobile
health care devices and the solution will be further scaled up to high end image scanning
systems such as CT scan, MRI etc. A futuristic system would be a health kiosk which will
scan the entire body and give automatic diagnosis.