A significant improvement by the early prototype of the MECoS approach could be shown immediately. It runs more efficiently than traditional systems and at the same time provides an equivalent learning performance and classification quality.
The required amount of processor cycles on the mobile device is remarkably below traditional, CS based approaches and lies at about 70% on average, i.e. allows savings of processor load on the mobile device of about 30%. It is interesting that in contrast to these savings, the MECoS approach produces a higher amount of data transfer between the mobile and the back-end system, experiments showed that the overall amount of data transfer is at up to 150% compared to traditional approaches.
On the other hand, the MECoS prototype clearly outperformed the traditional approaches when taking into account the number of data connections that were opened between mobile and back-end device. Here the MECoS approach allows savings of about 40 percent. It is important to take this into account when looking at the higher quantity of overall data transferred at 150%, because energy consumption characteristics determine the significance of these numbers and their impact on the overall results:
• Processor: energy intensive, but rather low when compared to communication efforts (GPRS data link)
• Data transmission: actively sending/receiving data: energy intensive, drains battery
• Opening up a data connection: most energy-intensive. When GPRS/UMTS data links are opened, so-called “bursts” are required to establish the connection ([208, ?]). These are short but extremely energy-intensive radio peaks of the cellphone or GSM modem. So, the impact of a higher amount of overall data transfer can be dropped by the remarkably lower number of data connections that have to be opened.
Remark on measured Communication Efforts
The MECoS prototype compressed the data for state/knowledge transfer with a simple RLE [172, 18, 35, 117] algorithm only. So, data transferred by the MECoS approach can be reduced massively by the application of a better compression algorithm. It demonstrates the improvement of the MECoS approach over traditional approaches also in terms of communication efforts: the outcome of the experiments has already shown that MECoS requires less communication efforts than traditional approaches in terms of the number of connections between mobile and backend device. When more sophisticated compression algorithms (e.g. LZW [152] or similar) are integrated into the MECoS approach, the required communication efforts can be reduced even more significantly. Even if the better compression uses more CPU cycles, CPU is cheaper than GPRS/UMTS - so this is a better deal in the end.
Final Verdict
If all three key energy consuming tasks are considered with their weights in combination, the advantage of the MECoS approach is clearly visible and it is proven that the previously set aims could be achieved - to recapitulate the stated aim of this research:
“The aim of this research project is to discover a novel approach based on Artificial Intelligence (AI) and Machine Learning (ML) methods, that results in a high quality and maximum energy- and computationally efficient operating Mobile Care System. This should enable a very economical, long period of autonomous operation without sacrificing the systems classification abilities / intelligence and result in applicable and cost-efficient mobile patient care systems.”
As the results of experiments show, the presented MECoS algorithm provides the same classification quality as other state-of-the-art AI-based monitoring systems by working much more resource and energy efficiently and therefore provides a significantly longer period of autonomous operation of the system - the research aim is clearly met.
Even though the development and advancement of mobile hardware is permanently progressing and devices work more efficiently and additionally also battery capacities are permanently increasing, the requirement remains - for a very efficient mode of operation of Mobile Care Systems for achieving a maximum autonomous system runtime - and derived from this - achieve cost savings. Faster processors and data transfer rates have an impact on numbers but the proportional superiority of the MECoS approach in
relation to traditional client-server based approaches will stay more or less unchanged: Mobile Care Systems that are based on the MECoS concept will always outperform traditional, client-server based systems because they inherently work more efficiently and therefore will always result in a longer autonomous system runtime.
A fundamental change in the results of this investigation would only occur if the proportions of the energy intensities for the performed tasks changed, i.e. required energy for processor, data communication and opening up data connections - in this case the results would have to be re-evaluated. But from the current point of view, this characteristic change is not to be expected in the near future.
Both measured systems, MECoS and also CS approaches, were based on simulation: in both cases the input data was generated by a developed software algorithm (both algorithms were fed with identical input data) and the classifiers were characteristic prototypes. To gather more detailed results, further experiments would be required, based on physiological input data (sensors measuring data from real patients) that are being measured in real world scenarios. This of course could result in (slight) changes of the outcome, but exceed the frame of this research project and it is very unlikely that the positive test of the stated hypothesis will be reverted.
Conclusion
8.1
Original Research Questions Addressed
A
S presented in the introduction of this thesis (see Chapter 1), the“How can we design a Mobile Care System that balances the conflicting needs of providing real-time complex distributed intelligence, using limited resources in terms of computational power and power requirements?”
The aim of this research project is to discover a novel approach based on Artificial Intelligence (AI) and Machine Learning (ML) methods, that results in a high quality and maximum energy- and computationally efficiently operating Mobile Care System. This should enable a very economical, long period of autonomous operation without sacrificing the systems classification abilities / intelligence and result in applicable and cost-efficient mobile patient care systems. Broader and increased usage of these systems will be the consequence.
After a detailed investigation of the state of the art in patient monitoring and artificial intelligence systems, a novel approach could be presented that addresses the above research questions, the so-called MECoS approach: The basic idea of the MECoS approach was to utilise modern wireless networks to overcome the limited computational resources of the mobile monitoring devices through integration of the assistance of powerful back-end systems with (nearly) unlimited power. The idea was, that the back-end systems should take over control in extraordinary cases - this should allow complex tasks and high classification quality by allowing the mobile monitoring devices to stay rather simple, portable and provide a very long autonomous system runtime. Technically, a heterogeneous distributed system was built, in which a classical online machine learning algorithm
was deployed. The speciality in this case is that the online machine learning algorithm works in a distributed way, i.e. the machine learning algorithm moves over multiple nodes within the system and integrates different available input parameters advancing a unique shared knowledge base on all nodes - so things learned on the back-end system are immediately available on the limited mobile monitoring device which would never have been able to learn these things itself.
A drawback of the MECoS approach is the rather high communicational efforts, as the knowledge base has to be transferred between the devices. It was unclear if this drawback would scatter the advantages of the MECoS approach, but as experiments have shown, data compression algorithms can be integrated and in the end the MECoS approach still offers significantly better results than the compared classical client server based approaches do.
8.2
Conclusion
The novel MECoS approach offers a significant advance in the application of a distributed machine learning algorithm that combines a battery-powered device with limited computational power to a remote and more powerful machine:
• Although running on a limited and portable device, through the distributed online machine learning algorithm the proposed system
achieves exactly the same classification quality as a powerful stationary computer system can provide.
• The overall processor cycles (i.e. energy usage and resource consumption) used on the mobile device per test run, shows MECoS clearly in front of classical CS (client server) approaches: MECoS is mainly operating at about 70% of CS’ consumption. Also at overall efforts for communication tasks MECoS outperforms CS with about 38% of CS’ consumption.
The presented MECoS approach represents a distributed online learning process that operates highly efficiently within an unbalanced distributed environment. Because there is only a single, shared knowledge base in combination with a moving classifier instance, a very dynamic and energy-efficient operation of the system can be achieved. The classifier provides high-quality results without the drawbacks of current Mobile Care Systems (“semi-portable” vs. “semi-intelligent” as mentioned above, see section 2.4).