Using Artificial Intelligence to Identify Patient Emotions across the Multiple Myeloma Diagnosis and Treatment Journey

Introduction: Multiple myeloma (MM) is a chronic hematological malignancy caused by the proliferation of malignant plasma cells in the bone marrow, and overproduction of monoclonal antibodies. Understanding the feelings of patients and caregivers about their lived experiences is central to supporting their quality of life, and artificial intelligence (AI) provides opportunities to enhance existing understanding, through language analytics.

Aims: To explore the capability of Natural Language Processing software (NLP) - a branch of AI - to extend understanding of lived experience, by identifying and classifying emotive words and phrases into categories. To explore patterns of emotions to identify whether participant's emotional experience varied at particular moments in the treatment journey.

Methods: The study involved qualitative interviews with n=24 MM patients and n=9 family caregivers in the UK, across three cohorts - i) those newly diagnosed (ND) within 12 months, ii) double-class exposed (DCE) and iii) triple-class refractory (TCR). Transcripts were manually coded by a team of researchers using Interpretative Phenomenological Analysis (IPA), identifying themes evidenced from 2541 excerpts. A patient journey code (e.g. pre-diagnosis, remission, relapse, etc.) was also tagged to each excerpt. This data was uploaded into AI software which analysed the excerpts and human-assigned codes, utilizing a pre-trained database of over 10,000 emotion words and phrases that were clustered into nine primary emotions and over 100 secondary emotions. Excerpts were tagged with primary and secondary emotion labels wherever an emotion match was found.

Results:AI software identified 10,313 words as emotion-related, from a total of 247,295 words (4.19 percent). Three findings were identified:

Joy, Trust, Anger and Fear made up 68 percent of all primary emotion identified. The distribution of emotional words across the nine primary emotion categories was concentrated into two positive emotions, Joy (17.7 percent) and Trust (14.5 percent) and two negative emotions, Fear (19.5 percent) and Anger (16.0 percent).

Secondary emotions identified more specificity of participant feelings. The most commonly identified secondary emotions expressing Joy were Power and Confidence and Optimism & Encouragement. Excerpts tagged with these leading secondary emotions provided context-specific examples of how emotion impacted participants either positively or negatively. For instance, secondary emotion impacts of Joy included i) feeling confident in treatment plans post diagnosis, ii) being encouraged by supportive clinical staff and iii) feeling empowered by improved energy levels achieved from starting treatment.

Frequency of certain emotions over-indexed at different points in the patient journey. By creating a baseline of the average incidence of emotion across all excerpts, the leading group of four primary emotions were plotted against the human-coded diagnostic journey steps to show peaks and troughs at different points of treatment. For example, Joy over-indexing peaked during remission and when accessing clinical trials for DCE and TCR cohorts, reflecting feelings of optimism and gratitude. Whereas Fear peaked at the stem cell transplantation stage for the ND cohort, evoking the anxiety and worry of isolation and treatment outcome. The highest peak of Anger was from TCR during remission and relapse, led by feelings of resentment towards the speed of returning symptoms and treatment regimes.

Conclusions: AI analytics can reveal sentiment patterns across multiple primary and secondary emotion types and patient journey phases by recognizing and categorizing patient and carer feelings. Therefore, AI can help clinical teams understand myeloma patients' changing emotions as they proceed through therapy. This enables identification of when patients and carers may need targeted psychological support, and the production/revision of support materials which are more specific rather than generic, by addressing emotional states at different stages of the disease. Nevertheless, due to patient profile variability and the number of participants in this qualitative study, future quantitative investigations are needed to assess the statistical significance of the AI findings.

Ali:Pfizer Ltd: Current Employment, Current equity holder in private company. Kuttschreuter:Pfizer Ltd: Divested equity in a private or publicly-traded company in the past 24 months, Ended employment in the past 24 months; Owkin: Current Employment, Current equity holder in private company. Wood:Pfizer Ltd: Current Employment, Current equity holder in private company. Harris:Pfizer Ltd: Consultancy, Research Funding.