Upinder S. Bhalla, Ph.D.
Professor & Dean, NCBS, Bengaluru, India
Watching the network change during the formation of associative memory
The process of learning is measured through behavioural changes, but it is of enormous interest to understand its cellular and network basis. We used 2-photon imaging of hippocampal CA1 pyramidal neuron activity in mice to monitor such changes during the acquisition of a trace conditioning task. One of the questions in such learning is how the network retains a trace of a brief conditioned stimulus (a sound), until the arrival of a delayed unconditioned stimulus (a puff of air to the eye). During learning, the mice learn to blink when the tone is presented, well before the arrival of the air puff.
The mice learnt this task in 20-50 trials. We observed that in this time-frame the cells in the network changed the time of their peak activity, such that their firing times tiled the interval between sound and air puff. Thus the cells seem to form a relay of activity. We also observed an evolution in functional connectivity in the network, as measured by groupings of correlated cells. These groupings were stable till the learning protocol commenced, and then changed. Thus we have been able to observe two aspects of network learning: changes in activity (relay firing), and changes in connectivity (correlation groups).
Srisairam Achuthan, Ph.D.
Senior Scientific Programmer, Research Informatics Division, Department of Information Sciences, City of Hope, CA, USA
Applying Machine learning for Automated Identification of Patient Cohorts
Srisairam Achuthan, Mike Chang, Ajay Shah, Joyce Niland
Patient cohorts for a clinical study are typically identified based on specific selection criteria. In most cases considerable time and effort are spent in finding the most relevant criteria that could potentially lead to a successful study. For complex diseases, this process can be more difficult and error prone since relevant features may not be easily identifiable. Additionally, the information captured in clinical notes is in non-coded text format. Our goal is to discover patterns within the coded and non-coded fields and thereby reveal complex relationships between clinical characteristics across different patients that would be difficult to accomplish manually. Towards this, we have applied machine learning techniques such as artificial neural networks and decision trees to determine patients sharing similar characteristics from available medical records. For this proof of concept study, we used coded and non-coded (i.e., clinical notes) patient data from a clinical database. Coded clinical information such as diagnoses, labs, medications and demographics recorded within the database were pooled together with non-coded information from clinical notes including, smoking status, life style (active / inactive) status derived from clinical notes. The non-coded textual information was identified and interpreted using a Natural Language Processing (NLP) tool I2E from Linguamatics.