Conclusion¶
In this investigation, we investigate the clustering of participants who conducted the Iowa gambling task (IGT). Engineered features included card deck choice, a health binary variable and the subject’s cumulative reward at different intervals (to monitor the balance between exploration and exploitation). In addition, we augment the existing data by adding participants from the Ahn et al. study. This enables us, to examine not only the behavioural differences between unhealthy and healthy individuals but also the varying decision-making deficiencies between opioid and stimulant dependent individuals.
Our most significant results are the following:
Exploratory data analysis of the various studies contradicts the general assumption of IGT, i.e., the preference of healthy individuals to seek long-term reward. Rather than picking the two advantageous decks (C, D), ad-hoc analysis demonstrated participants generally prefer one of the advantageous (deck D) and one of the disadvantageous (deck B). Possible reasons for this observed discrepancy may be found in the particular payoff scheme of the study and the resulting inter-study biases. Alternatively, we hypothesis that healthy individuals are influenced by both long term reward and immediate gain/loss frequency.
Similarly, we observed a high inter-study and inter-individual variability in IGT performance in healthy participants. Participant variability could be due to divergent psychological attributes of the healthy participants such as learn behaviour, a propensity to gambling, impulsivity or different decision-making strategies. In addition, different IGT versions may explain the inter-study discrepancy.
Although both heroin and amphetamine users display poor decision making, the average heroin user displays a poorer take-home reward (approx. $200 difference). Different classes of drugs might have different effects on decision-making behaviour. As mentioned previously, pre-clinical trails concluded that stimulant and opiate users display different behavioural effects. Stimulants tend to produce arousing and activating effects. In contrast, opiates produce mixed inhibitory and excitatory effects.
K-means could reasonably group all unhealthy individuals into a single cluster, but healthy individuals were distributed evenly. K-means performed poorly when we examined the clusters by study.
The devised federated k-means algorithm resulted in a 10% increase in the sum square error.
Future work¶
There are several possibilities to extend this work in future. Currently, both heroin and amphetamine addicts are grouped into the same cluster using K-means (even with different numbers of K and principal components). We plan to experiment with hierarchical clustering algorithms that may be able to model the distinction between these subgroups in a wider ‘poor decision making’ cluster. Furthermore, sub-groups of healthy participants may be revealed with associated advantageous or disadvantageous decision behaviour. In our investigation, we analysed our clusters by study and found little correspondence. However, we would like to analyse our clusters by payoff scheme in an attempt to investigate how participants decision making abilities are impacted by the type of IGT performed. In addition, we plan to train a reinforcement learning model on the datasets and perform clustering utilizing the parameters of that model. Similar endeavours have shown to be fruitful, with such parameters often increasing the interoperability of results. Our devised federated k-means algorithm could also be improved by incorporating mini-batch k-means. This would result in only a trivial reduction in accuracy and would be particularly useful if the decentralized server was a mobile phone. Finally, we are interested in incorporating other features about the subjects such as socio-economic status, gender, and a chronic gambling addiction indicator. We hope such features might uncover specific card decision patterns or behavioural inabilities during the task.