**5. Discussion and conclusions**

The results of our tests show great promise for how these approaches to machine learning use in precision medicine can be used, while nonetheless highlighting some significant shortcomings which will need to be considered should this framework become common practice. Most notably, we see that model's which over-fit the available data are extremely detrimental to the system, even if the underlying model would be better with all data immediately available to it. This is shown very clearly with the effectiveness of pruning in all our models, with clear gains in accuracy observed, likely as a result of the process helping counteract over-fitting resulting from over-parameterized models. Finding an "ideal" model for a given task is already a difficult task, and our results show that this is only exacerbated by the conditions of a clinical environment. Nevertheless, there is clearly potential in this framework, with the Convolutional network tested on clinical-like batch data being near identical in effectiveness to its static counterpart trained on the full dataset.

We are also optimistic that many opportunities remain for improvement in progressive learning implementations. Our current implementation of the progressive learning framework is locked to a specific set of initial data inputs, being unable to add new ones should they become available. In theory, this could be as simple as adding a new set of initial blocks to the existing network, in effect acting like a "progression" stage with custom new blocks (as well as an update to existing new block generation procedures to match). However, this has a number of issues that we have not, at present, found a way to resolve. First, each branch is likely to "learn" at different rates, resulting in one set of blocks associated with a given set of input data containing more redundant features per-progression stage than the rest. This proves problematic during pruning, however; we either over-prune blocks with important features within them, or under-prune those which contain an abundance of redundant and/or noisecontributing features. We believe this can be resolved, but were simply unable to do so by the time of this publication.

Another potential improvement would be to "carry-over" output layers weights between progression stages. This would allow for the network to have better forward transfer, so long as the task's end goal (categorical prediction, single metric estimation etc.) remains the same. In our implementation, this is currently not the case, with the output layer being regenerated every cycle, keeping it in line with the original Progressive Learning framework's design [29]. The difficulty of implementing such as system, as well as its effectiveness in improving model outcomes, has yet to be tested.

One other major hurdle is that of long term memory cost. As currently implemented, pruning does not actually remove parameters from the model; it simply masks them out during training and evaluation, preventing them from contributing to the models predictions. While this improves the speed and accuracy of the model being generated, its memory footprint expands infinitely as more cycles are run. Resolving this issue is difficult, however, requiring the model to effectively fully re-construct itself to account for any now-removed parameters. Doing so would allow the model to come to a "static" memory cost, as the number of pruned parameters approaches the number of new ones added every cycle. In turn, this would enable applications where the model is expected to existing for very long duration in limited resource systems. Such compression techniques are an ongoing field of research at time of writing; as such, we believe such a implementation will be possible in the near future.

Finally, testing our methodology on a real-world clinical dataset is needed before we can be sure it is truly effective. While the CIFAR-10 dataset [34] has been shown to work effectively for machine learning testing purposes, our assumptions about clinical data still need to be confirmed. We intend to put our framework to the test soon, assessing its effectiveness at predicting DCM severity using the DCM data mentioned throughout this chapter; nevertheless, this framework should be considered experimental until such results (from ourselves or others) are acquired. Continual learning systems trained for clinical data also retain the limitations of continual learning, such as increased potential to over-fit and the inability to transfer new knowledge obtained to help with the understanding of prior knowledge. Modifications to the progression procedure have been proposed to amend this [29], though these have not been tested at time of writing.

Overall, however, we believe our framework for machine learning system design in precision medicine should work well as a road-map for future research, even

*Delivering Precision Medicine to Patients with Spinal Cord Disorders; Insights into… DOI: http://dx.doi.org/10.5772/intechopen.98713*

though refinements remain to be made. With systems such as the Progressive Learning framework available, these new systems can adapt to changes in data trends while accepting new data in effectively random batches, both important requirements for a clinical environment. Well designed data storage and management also allows such systems to easily access, update, and report important metrics to all necessary parties, while remaining open to changes as new research is completed. Through the application of these techniques, modern medicine should be able to not only adapt to the age of information, but to benefit immensely from it.
