In some circles, you hear that we need new better quality measures in oncology, but it's not always clear how to go about discovering them.
In this exercise, we try to use "guided creativity" to see if AI can help (Chat GPT). Since good measures begin with the patient, we asked Chat GPT to list 15 problems that cancer patients face. THEN, list an intervention or remedy for each, if possible. THEN, take each problem-solution pair, and sketch it as a "quality measure." This is essentially guided brainstorming for the AI, and some of the 15 may be hopeless mistakes, others obvious, but perhaps a few could be "promising."
##
Oncology needs a new generation of quality measures that better capture what matters to patients. How might we discover them? In this exercise, we guided a large language model into "creative brainstorming" through three steps:
- Identify 15 important problems patients face during cancer care,
- propose an achievable intervention for each, and
- suggest measurements that could show whether care improved.
The table below presents the resulting candidates, graded and ranked by their promise as actionable quality measures. They require further refinement and validation. Some could benefit from AI-assisted data collection; others need no AI at all. All begin with a practical definition of quality: the right service, at the right time, helping the patient.
##
| Patient problem | Reasonable, achievable intervention | Measures that would track improvement |
|---|---|---|
| 1. Pain, nausea, fatigue, insomnia, or other burdensome symptoms persist between visits. A. | Collect brief symptom reports; assign alerts to a responsible clinician; treat and reassess, with palliative support when needed. | Days with severe symptoms; percentage of significant reports receiving timely action; percentage achieving meaningful symptom improvement after intervention. |
| 2. Severe treatment toxicity develops without adequate preparation, prevention, or rapid help. A. | Provide regimen-specific prevention, warning-sign education, accessible symptom reporting, and urgent clinical triage. | Percentage receiving indicated preventive measures; time from red-flag report to clinical action; severe toxicity and related emergency visits, adjusted for regimen and risk. |
| 3. After serious toxicity, treatment continues without reconsidering whether its benefits justify its burdens. A. | Trigger a renewed discussion of prognosis, alternatives, dose changes, treatment breaks, stopping, and the patient’s priorities. | Percentage of triggering events followed by reassessment before the next decision; patient understanding of options; patient-reported alignment of the revised plan with priorities. |
| 4. A useful treatment option is missed because appropriate biomarker testing was omitted, delayed, or overlooked. A. | Identify testing indications, track results, and review actionable findings before the relevant treatment decision. | Percentage with indicated results available when needed; percentage of actionable findings addressed through treatment, referral, or an explained alternative. |
| 5. Insurance denials, authorization errors, and billing disputes consume time and interrupt care. B. | Assign staff ownership of authorizations and appeals; proactively correct errors and update patients. | Days of care delayed by insurance; unresolved disputes; patient hours spent managing administrative problems; percentage resolved before care is interrupted. |
| 6. Transportation, work, childcare, or caregiving obligations make the care plan impractical. B. | Arrange practical support; consolidate visits; use local care or telehealth when appropriate. | Missed or delayed care attributable to practical barriers; percentage of identified barriers resolved; patient travel, waiting, and administrative hours per treatment month. |
| 7. Patient does not understand treatment intent, likely benefit, alternatives, or burdens. B. | Plain-language decision discussion, interpreter when needed, and teach-back; elicit the patient’s priorities. | Percentage correctly understanding treatment intent and major tradeoffs; patient-reported involvement in decisions; unresolved decisional conflict. |
| 8. Cannot find an appropriate specialist in network with a timely appointment. B. | Navigator arranges an appropriate appointment; escalates network inadequacy and urgent referrals. | Days from referral to completed consultation; percentage seen within an urgency-specific target; percentage unable to obtain care. |
| 9. Oral cancer treatment is taken incorrectly or interrupted because of confusion, interactions, or refill problems. C. | Pharmacist review, teach-back, early follow-up, refill tracking, and explicit instructions for treatment holds. | Medication discrepancies resolved; days without medication because of access failures; dosing errors; concordance with the current agreed treatment plan. |
| 10. Anxiety, depression, isolation, or caregiver exhaustion goes unaddressed. C. | Brief assessment followed by accessible counseling, social work, peer support, or caregiver assistance. | Percentage of identified needs receiving an accepted service; time to first service; change in distress or caregiver burden; unresolved requests for help. |
| 11. Neither clinician nor patient raises dying or hospice until a crisis removes meaningful choices. C. | Offer timely prognosis and goals discussions; involve palliative care; explain and arrange hospice when appropriate and desired. | Percentage offered discussion while able to participate; time from hospice request to access; patient/family-reported concordance between desired and delivered end-of-life care. |
| 12. Treatment costs threaten medication access, housing, food, or family finances. C. | Provide prospective cost estimates, financial navigation, assistance applications, and discussion of clinically suitable affordable options. | Change in financial distress score; percentage delaying or forgoing care because of cost; percentage of identified financial barriers resolved. |
| 13. Treatment erodes mobility, nutrition, independence, or the ability to pursue valued activities. D. | Assess function and nutrition; provide rehabilitation and dietary support; adapt treatment when appropriate. | Change in function and nutritional status; percentage receiving indicated support; percentage maintaining or recovering a patient-selected activity or goal. |
| 14. Diagnosis or staging remains uncertain, delaying or misdirecting treatment. D. | Track outstanding pathology and staging; obtain expert review when uncertainty could change management. | Days to a treatment-ready diagnosis; percentage of consequential diagnostic uncertainties resolved before treatment; repeat procedures caused by inadequate workup. |
| 15. Fragmented care leaves the patient carrying information between teams or managing conflicting instructions. D. | Name an accountable coordinator; maintain a shared plan; reconcile medications and close referral and discharge handoffs. | Percentage of handoffs completed with needed information; unresolved conflicting instructions; patient-reported coordination failures; duplicative tests attributable to fragmentation. |
Table 1. After the table was generated, we asked AI to return to the table and grade each row, then re-order them. The A’s have the clearest links between a recognizable need, a specific action, and assessable benefit. The B’s are also strong ideas but requiring clearer attribution or interpretation. The C’s face greater challenges in defining eligibility, capturing actual experience, or separating care effects from other influences. The D’s bundle heterogeneous problems or risk rewarding speed and documentation without establishing benefit. 12/13 might improve substantially by selecting one functional domain and a defined patient group.
In Closing
These are proposed measurement concepts, rather than fully specified, validated measures. Percentages would use patients with the relevant need as their denominator, and timeliness standards would reflect clinical urgency. The central question throughout is whether an identified patient problem received an effective response.
##
Discuss in the style of a JAMA OP ED that comments on a research paper. Lessons learned?
##
Finding Better Quality Measures: Can AI Work Backward From Patient Needs?
Oncology needs better quality measures, but the call for a new generation of measures leaves an important question unanswered: How do we discover them? AI expands the information available to us. Turning that information into meaningful measures of care requires a clear account of what patients need and what health care organizations can reasonably do to help.
We explored a simple approach using a guided large language model. First, identify important problems patients experience during cancer care. Second, propose a feasible intervention for each. Third, describe a measurement that could show whether care improved. This sequence makes the patient’s problem the starting point and requires a plausible connection between measurement and benefit. The resulting fifteen-row table is a set of hypotheses about what better quality measurement might look like.
Several proposals deserve particular attention. Persistent symptoms offer a direct connection between need, action, and benefit. A patient reports substantial pain or nausea; the team responds; subsequent reports show whether the intervention helped. Measuring both responsiveness and relief brings us closer to the experience we want to improve. Insurance barriers provide another concrete opportunity. Days of treatment delayed by an unresolved authorization, or hours a patient spends correcting administrative errors, describe burdens that can be consequential even when the eventual treatment is clinically appropriate. These measures would require accountability shared appropriately among practices, hospitals, and insurers.
Perhaps the most interesting proposal concerns reassessment after serious toxicity. A treatment decision that was reasonable several months ago may deserve reconsideration today. Quality includes recognizing that turning point, explaining the changing tradeoffs, and revisiting the patient’s priorities. AI could help identify such moments across clinical notes, laboratory findings, and patient messages. A successful measure would recognize thoughtful decisions to continue, modify, pause, or stop treatment.
The exercise also exposed weak proposals. “Resolving diagnostic uncertainty” combines heterogeneous situations and could reward speed without establishing accuracy. “Improving coordination” can readily deteriorate into counting completed handoffs. Both concerns matter, but they need sharper definitions before becoming useful measures. This critical review is part of the method: generating candidates should be followed by selection, refinement, and testing.
The distinction between discovering clinical knowledge and measuring care also becomes clearer. An analysis of thousands of records might reveal differences in treatment response between patient groups. Such findings can inform care once adequately validated. A quality measure asks a further question: Did this patient receive an appropriate service, at an appropriate time, with an informed choice and effective follow-through?
For the authors' panel at a policy conference next week, the exercise illustrates a practical role for AI in health policy: helping convert a broad aspiration into concrete proposals that people can examine and improve. Some proposals require AI to become feasible; others need straightforward data collection and organizational responsibility. The next step is to select a few promising candidates and test whether using them leads to better patient experiences and outcomes. Better measures earn their place by helping care become better.
##