REFORMS: Consensus-based Recommendations for Machine-learning-based Science.

Machine learning (ML) methods are proliferating in scientific research. However, the adoption of these methods has been accompanied by failures of validity, reproducibility, and generalizability. These failures can hinder scientific progress, lead to false consensus around invalid claims, and undermine the credibility of ML-based science. ML methods are often applied and fail in similar ways across disciplines. Motivated by this observation, our goal is to provide clear recommendations for conducting and reporting ML-based science. Drawing from an extensive review of past literature, we present the REFORMS checklist (recommendations for machine-learning-based science). It consists of 32 questions and a paired set of guidelines. REFORMS was developed on the basis of a consensus of 19 researchers across computer science, data science, mathematics, social sciences, and biomedical sciences. REFORMS can serve as a resource for researchers when designing and implementing a study, for referees when reviewing papers, and for journals when enforcing standards for transparency and reproducibility.

Evaluation of Recurrent Neural Network Models for Parkinson's Disease Classification Using Drawing Data.

Parkinson's disease is a disorder that affects the neurons in the human brain. The various symptoms include slowness of motor functions (bradykinesia), motor instability, speech impairment and in some cases, psychiatric effects such as hallucinations. Most of these, however, are also common side effects of natural aging. This makes an accurate diagnosis of Parkinson's disease a challenging task. Some breakthroughs have been made in recent years with the help of deep learning. This work aims at considering figure drawing data as a time series of coordinates, angles and pressure readings to train recurrent neural network models. In addition, the work compares two recurrent network models, Long Short-Term Memory and Echo State Networks, to explore the advantages and disadvantages of both architectures.

Significant cognitive decline in Parkinson's disease exacerbates the reliance on visual feedback during upper limb reaches.

While upper limb reaches are often made in a feed-forward manner, visual feedback during the movement can be used to guide the reaching hand towards a target. In Parkinson's disease (PD), there is evidence that the utilisation of this visual feedback is increased. However, it is unclear if this is due solely to the characteristic slowness of movements in PD providing more opportunity for incorporating visual feedback to modify reach trajectories, or whether it is due to cognitive decline impacting (feed-forward) movement planning ability. To investigate this, we compared reaction times and movement times of reaches to a target in groups of PD patients with normal cognition (PD-NC), mild cognitive impairment (PD-MCI) or dementia (PD-D), to that of controls with normal cognition (CON-NC) or mild cognitive impairment (CON-MCI). Reaches were undertaken with full visual feedback (at a 'natural' and 'fast-as-possible' pace); with reduced visual feedback of the reaching limb to an illuminated target; and without any visual feedback to a remembered target with eyes closed. The PD-D group exhibited slower reaction times than all other groups across conditions, indicative of less efficient movement planning. When reaching to a remembered target with eyes closed, all PD groups exhibited slower movement times relative to their natural pace with full visual feedback. Crucially, this relative slowing was most pronounced for the PD-D group, compared to the PD-MCI and PD-NC groups, suggesting that substantial cognitive decline in PD exacerbates dependence on visual feedback during upper limb reaches.

Machine learning discriminates a movement disorder in a zebrafish model of Parkinson's disease.

Animal models of human disease provide an in vivo system that can reveal molecular mechanisms by which mutations cause pathology, and, moreover, have the potential to provide a valuable tool for drug development. Here, we have developed a zebrafish model of Parkinson's disease (PD) together with a novel method to screen for movement disorders in adult fish, pioneering a more efficient drug-testing route. Mutation of the PARK7 gene (which encodes DJ-1) is known to cause monogenic autosomal recessive PD in humans, and, using CRISPR/Cas9 gene editing, we generated a Dj-1 loss-of-function zebrafish with molecular hallmarks of PD. To establish whether there is a human-relevant parkinsonian phenotype in our model, we adapted proven tools used to diagnose PD in clinics and developed a novel and unbiased computational method to classify movement disorders in adult zebrafish. Using high-resolution video capture and machine learning, we extracted novel features of movement from continuous data streams and used an evolutionary algorithm to classify parkinsonian fish. This method will be widely applicable for assessing zebrafish models of human motor diseases and provide a valuable asset for the therapeutics pipeline. In addition, interrogation of RNA-seq data indicate metabolic reprogramming of brains in the absence of Dj-1, adding to growing evidence that disruption of bioenergetics is a key feature of neurodegeneration.This article has an associated First Person interview with the first author of the paper.

Using echo state networks for classification: A case study in Parkinson's disease diagnosis.

Despite having notable advantages over established machine learning methods for time series analysis, reservoir computing methods, such as echo state networks (ESNs), have yet to be widely used for practical data mining applications. In this paper, we address this deficit with a case study that demonstrates how ESNs can be trained to predict disease labels when stimulated with movement data. Since there has been relatively little prior research into using ESNs for classification, we also consider a number of different approaches for realising input-output mappings. Our results show that ESNs can carry out effective classification and are competitive with existing approaches that have significantly longer training times, in addition to performing similarly with models employing conventional feature extraction strategies that require expert domain knowledge. This suggests that ESNs may prove beneficial in situations where predictive models must be trained rapidly and without the benefit of domain knowledge, for example on high-dimensional data produced by wearable medical technologies. This application area is emphasized with a case study of Parkinson's disease patients who have been recorded by wearable sensors while performing basic movement tasks.

A New Evolutionary Algorithm-Based Home Monitoring Device for Parkinson's Dyskinesia.

Parkinson's disease (PD) is a neurodegenerative movement disorder. Although there is no cure, symptomatic treatments are available and can significantly improve quality of life. The motor, or movement, features of PD are caused by reduced production of the neurotransmitter dopamine. Dopamine deficiency is most often treated using dopamine replacement therapy. However, this therapy can itself lead to further motor abnormalities referred to as dyskinesia. Dyskinesia consists of involuntary jerking movements and muscle spasms, which can often be violent. To minimise dyskinesia, it is necessary to accurately titrate the amount of medication given and monitor a patient's movements. In this paper, we describe a new home monitoring device that allows dyskinesia to be measured as a patient goes about their daily activities, providing information that can assist clinicians when making changes to medication regimens. The device uses a predictive model of dyskinesia that was trained by an evolutionary algorithm, and achieves AUC>0.9 when discriminating clinically significant dyskinesia.

Artificial Epigenetic Networks: Automatic Decomposition of Dynamical Control Tasks Using Topological Self-Modification.

This paper describes the artificial epigenetic network, a recurrent connectionist architecture that is able to dynamically modify its topology in order to automatically decompose and solve dynamical problems. The approach is motivated by the behavior of gene regulatory networks, particularly the epigenetic process of chromatin remodeling that leads to topological change and which underlies the differentiation of cells within complex biological organisms. We expected this approach to be useful in situations where there is a need to switch between different dynamical behaviors, and do so in a sensitive and robust manner in the absence of a priori information about problem structure. This hypothesis was tested using a series of dynamical control tasks, each requiring solutions that could express different dynamical behaviors at different stages within the task. In each case, the addition of topological self-modification was shown to improve the performance and robustness of controllers. We believe this is due to the ability of topological changes to stabilize attractors, promoting stability within a dynamical regime while allowing rapid switching between different regimes. Post hoc analysis of the controllers also demonstrated how the partitioning of the networks could provide new insights into problem structure.

Using epigenetic networks for the analysis of movement associated with levodopa therapy for Parkinson's disease.

Levodopa is a drug that is commonly used to treat movement disorders associated with Parkinson's disease. Its dosage requires careful monitoring, since the required amount changes over time, and excess dosage can lead to muscle spasms known as levodopa-induced dyskinesia. In this work, we investigate the potential for using epiNet, a novel artificial gene regulatory network, as a classifier for monitoring accelerometry time series data collected from patients undergoing levodopa therapy. We also consider how dynamical analysis of epiNet classifiers and their transitions between different states can highlight clinically useful information which is not available through more conventional data mining techniques. The results show that epiNet is capable of discriminating between different movement patterns which are indicative of either insufficient or excessive levodopa.

Computational approaches for understanding the diagnosis and treatment of Parkinson's disease.

This study describes how the application of evolutionary algorithms (EAs) can be used to study motor function in humans with Parkinson's disease (PD) and in animal models of PD. Human data is obtained using commercially available sensors via a range of non-invasive procedures that follow conventional clinical practice. EAs can then be used to classify human data for a range of uses, including diagnosis and disease monitoring. New results are presented that demonstrate how EAs can also be used to classify fruit flies with and without genetic mutations that cause Parkinson's by using measurements of the proboscis extension reflex. The case is made for a computational approach that can be applied across human and animal studies of PD and lays the way for evaluation of existing and new drug therapies in a truly objective way.

The incorporation of epigenetics in artificial gene regulatory networks.

Artificial gene regulatory networks are computational models that draw inspiration from biological networks of gene regulation. Since their inception they have been used to infer knowledge about gene regulation and as methods of computation. These computational models have been shown to possess properties typically found in the biological world, such as robustness and self organisation. Recently, it has become apparent that epigenetic mechanisms play an important role in gene regulation. This paper describes a new model, the Artificial Epigenetic Regulatory Network (AERN) which builds upon existing models by adding an epigenetic control layer. Our results demonstrate that AERNs are more adept at controlling multiple opposing trajectories when applied to a chaos control task within a conservative dynamical system, suggesting that AERNs are an interesting area for further investigation.

Computational models of signalling networks for non-linear control.

Artificial signalling networks (ASNs) are a computational approach inspired by the signalling processes inside cells that decode outside environmental information. Using evolutionary algorithms to induce complex behaviours, we show how chaotic dynamics in a conservative dynamical system can be controlled. Such dynamics are of particular interest as they mimic the inherent complexity of non-linear physical systems in the real world. Considering the main biological interpretations of cellular signalling, in which complex behaviours and robust cellular responses emerge from the interaction of multiple pathways, we introduce two ASN representations: a stand-alone ASN and a coupled ASN. In particular we note how sophisticated cellular communication mechanisms can lead to effective controllers, where complicated problems can be divided into smaller and independent tasks.

Characterising neurological time series data using biologically motivated networks of coupled discrete maps.

Artificial biochemical networks (ABNs) are a class of computational dynamical system whose architectures are motivated by the organisation of genetic and metabolic networks in biological cells. Using evolutionary algorithms to search for networks with diagnostic potential, we demonstrate how ABNs can be used to carry out classification when stimulated with time series data collected from human subjects with and without Parkinson's disease. Artificial metabolic networks, composed of coupled discrete maps, offer the best recognition of Parkinsonian behaviour, achieving accuracies in the region of 90%. This is comparable to the diagnostic accuracies found in clinical diagnosis, and is significantly higher than those found in primary and non-expert secondary care. We also illustrate how an evolved classifier is able to recognise diverse features of Parkinsonian behaviour and, using perturbation analysis, show that the evolved classifiers have interesting computational behaviours.

Potential for Raman spectroscopy to provide cancer screening using a peripheral blood sample.

Cancer poses a massive health burden with incidence rates expected to double globally over the next decade. In the United Kingdom screening programmes exists for cervical, breast, and colorectal cancer. The ability to screen individuals for solid malignant tumours using only a peripheral blood sample would revolutionise cancer services and permit early diagnosis and intervention. Raman spectroscopy interrogates native biochemistry through the interaction of light with matter, producing a high definition biochemical 'fingerprint' of the target material. This paper explores the possibility of using Raman spectroscopy to discriminate between cancer and non-cancer patients through a peripheral blood sample. Forty blood samples were obtained from patients with Head and Neck cancer and patients with respiratory illnesses to act as a positive control. Raman spectroscopy was carried out on all samples with the resulting spectra being used to build a classifier in order to distinguish between the cancer and respiratory patients' spectra; firstly using principal component analysis (PCA)/linear discriminant analysis (LDA), and secondly with a genetic evolutionary algorithm. The PCA/LDA classifier gave a 65% sensitivity and specificity for discrimination between the cancer and respiratory groups. A sensitivity score of 75% with a specificity of 75% was achieved with a 'trained' evolutionary algorithm. In conclusion this preliminary study has demonstrated the feasibility of using Raman spectroscopy in cancer screening and diagnostics of solid tumours through a peripheral blood sample. Further work needs to be carried out for this technique to be implemented in the clinical setting.

Modelling biological evolvability: implicit context and variation filtering in enzyme genetic programming.

This paper describes recent insights into the role of implicit context within the representations of evolving artefacts and specifically within the program representation used by enzyme genetic programming. Implicit context occurs within self-organising systems where a component's connectivity is both determined implicitly by its own definition and is specified in terms of the behavioural context of other components. This paper argues that implicit context is an important source of evolvability and presents experimental evidence that supports this assertion. In particular, it introduces the notion of variation filtering, suggesting that the use of implicit context within representations leads to meaningful variation filtering whereby inappropriate change is ignored and meaningful change is encouraged during evolution.