Mobile Health Application Development Considerations
Execute the network on the server, with the RestFUL request containing the authentication token and input tensor and the RestFUL response containing the output tensor and a user-comprehensible exception list.
Execute the network on the client, which will require implementing a way to transfer learned parameters to the custom Java and objective-C containers for execution on the Android and iOS clients respectively.
Network training must be on the server and leverage C, C++, or nvcc compiled code, which can also be accessed through Java or Python wrapper libraries.
Mental Health Considerations
Mitigating depression, which in non-clinical terms, is a lack of hope. Clinically, it is usually a genetic vulnerability expressed in the wake of an affecting sequence of some combination of developmental and recent events. It is almost always a complex of interacting factors.
[Wish] to be able to import all the relevant information: key words, text, images, videos from Facebook, Instagram, Twitter to prevent or report the existence of depression and correctly interpret the nature, intensity and frequency of depression symptoms. ... a patient profile ...
This gathering is subject to HIPPA law in the United States, and there is similar law in other jurisdictions, so there will need to be a waiver process integrated in to the sign-up and use of the application. You will need to consult experienced legal talent for both legalese and the requirements regarding integration.
Analysis of the Proposal
There are several inputs mentioned.
- Social networking content ++
- Patient profile ++
- Camera (for facial recognition and ambient lighting)
- Sensors that can measure changes in mood (not clear what properties would be measured)
- Device lock time (not clear what this means or why it helps)
- GPS sensor ++
- Recent calls ++
There are several models mentioned.
- Facial recognition (probably employing CNN, along with semi-automatic user assisted panning and ranging to get a good image for recognition)
- Model of cognitions, such as doubt, doom, and gloom and their converse confidence, hope, and excitement
- Sleep patterns
- Increased isolation ++
- Map of depressing versus encouraging locations (since the GPS is an input) ++
- Map of depressing versus encouraging contacts ++
- Map of depressing versus encouraging thought patterns ++
There are several outputs mentioned.
- Virtual LED lights that react to emotions
- Educational resources
- Medication reminder
- Meeting reminder ++
There are several pages mentioned, which would contain both inputs and outputs.
- Achievement board
- Dream chart
- To-do list
- Thinking patterns ++
- Daily thoughts ++
- Techniques and activities ++
- Counselling scheduler ++
- Mental health chatbot
Results of Analysis
These pages may be useful but were not explicitly mentioned.
- Medication check-off ++
- Proximity stop-go lights ++
- Contact block recommendations ++
The full set of requirements in the question would require a large project scope, probably requiring years of development and involving a sizable team.
- An excellent systems architect
- AI engineers
- General mobile developers
- A data base manager
- Middle tier programmers
- A high level clinician
- A product manager
- A good project manager
- Team leaders
- Marketing specialists
- Web designers
- A security specialist
- Access to a health care legal expert
- Technically savvy senior management
It is feasible to succeed in this domain, but there are already many mobile health care companies that may compete in this particular space. A market advantage will be necessary, which is likely either a very large sum of investment capital or a particularly innovative and efficient approach with an only somewhat large sum of investment capital.
To gain financial traction for a project like this, a set of page markups will not suffice. A more convincing proof of concept will likely be necessary. Consider beginning in phases, starting with ideas that will differentiate the product from others and that are easiest to implement. See the double plus signs above.
The highest degree of difficulty are the chatbot, the to-do list, and the cognition modelling (CBT) capabilities. Avoid placing those in your project scope at the beginning. The ideas are not new, and no one has, with millions in their budget, accomplished much that meets the minimum standards of effectiveness yet.
Don't use a cloud service unless you do your own strong encryption. You don't want legal exposure because someone hacked the cloud security layers and gained access to medical data. There are teams of people who go to work every day just to hack those layers, and some of the attacks succeed nearly every week.
Medical mobile applications are highly proprietary. There are no open source solutions and eBooks will not exist this early in the development of this space because the expertise is narrow and in demand. Those who know how to do it well are constrained by nondisclosure agreements, which is why I can't go into further detail down specific roads.
Expanding on Analysis
These are some thoughts that are not constrained by legal agreements.
The first principle to understand when developing a new product or service is that the simplest models are ones where the correlation between easy to acquire data and various states can be modeled without much computing power. Also, health care related mobile apps will not be used if they require much writing. Everything must be quick and easy, or the usage patterns will suffer. There must also be accountability and incentive, which is part of the creative design of the product or service.
This is one example of such a system, which can be implemented with relative ease compared to the rather large project scoped in the question.
- Clinicians can sign up and then invite their patients to sign up.
- Patients can sign up and agree to the legalese prepared by legal talent to comply with medical record legislation.
- clinicians can write, at the end of sessions, suggestions to the patient.
- Patients can interact with the suggested items in extremely simple and quick ways.
- The clinicians can label the data, meaning that they can assign scores in specific categories to what the patients entered.
- The AI, invisible to all users, then learns what scores are assigned to what kinds of entries so that, at a superficial level, the clinician's expertise as specifically applies to that one patient is trained into an artificial network.
- Iconic rewards and warnings can then be sent to the patient as a real time guidance device.
- Proximity stop-go lights can be lit in addition to rewards and warnings. As a patient approaches a location or begins to contact a person that is either helpful or a hindrance, the color of the light indicates whether to go ahead, approach with caution, or stop.
- Trends across patients and clinicians can be learned so that various locations, contacts, and other easy to identify specifics can be learned as helpful or dangerous without further clinician entries.
Iconic rewards and warnings and stop-go lights are more directly functional than mood lights because mood as a consequence of action is easy to model. Although actions may stem from moods, the measurement of mood poses practical difficulty and is therefore much more difficult to model in a mobile application system, even with the best AI has to offer today.
Very little of this is likely done yet, and to do it well might be highly effective and a market differentiator. It is not patentable because it does not meet the burden of patentability. Specifically a typical solution that any of those experienced in the field would arrive at using normal methods in that field cannot be considered a new invention. Therefore, the first person who reads this and implements may gain the market advantage.
The reason that a medication check-off is better than merely a reminder is because people may or may not heed the reminder immediately for practical reasons, so checking off that the medication was taken is key, just in the same way a week long pill box works.
When a particular contact leads to signs of depression, an LSTM network or a GRU network can learn those correlations and make contact block recommendations, since most social networks, text messaging clients, and phone services provide contact blocking.
Commit to Learning Key Technology Without Delay
To learn about AI, there are several Q&A pages here.