Evaluation of ChatGPT’s Potential in Tailoring Gynecological Cancer Therapies

Results

A total of 16 patients with gynecological carcinomas were identified, including five each with ovarian and endometrial carcinoma and six with cervical carcinoma. The age of the patients examined ranged from 27 to 86 years. The study included a comprehensive examination of both early-stage carcinomas and advanced tumors [Fédération Internationale de Gynécologie et d’Obstétrique (FIGO) IA-IVB]. Notably, the endometrial carcinoma group primarily consisted of early-stage carcinomas, while the ovarian carcinoma group was characterized by a higher proportion of advanced tumors. ChatGPT was able to successfully design four treatment plans for each patient.

Ovarian carcinoma. Answer scores for ovarian carcinoma patients. Across all queries for ovarian carcinoma (OC), ChatGPT achieved an average score of 0.75 points according to the answer scoring system [range: −1 points (minimum) to 2 points (maximum)]. For patients 1-4, 11 out of 16 treatment plans (11 out of 16 AI-Rs) were awarded one point, as the proposed treatments were correct and described in detail, but individual treatment steps (e.g., genetic testing, further surgical interventions or targeted therapy with VEGF and PARP inhibitors) were missing from the treatment plan (see Table I). For patient 5, one point was deducted for each AI-R for the mention of chemotherapy, although the patient’s general condition was limited. The score of 0 points was awarded a total of five times among all treatment plans for ovarian cancer (5 out of 20 AI-Rs) because, in addition to detailed, though incomplete therapy recommendations, incorrect treatment modalities such as radiotherapy were also suggested by ChatGPT (see Table I). The maximum score was not achieved once in the OC subgroup.

Surgical treatment. In the present study, the ITB identified an indication for surgical treatment in all of the OC patients. With the exception of one patient, the recommendations by ChatGPT were consistent with this. Patient 4, a 45-year-old perimenopausal woman with a low-grade serous carcinoma of the right ovary, contacted the clinic for a second opinion after having undergone hysterectomy, salpingectomy on both sides, biopsies of the ovaries and omentectomy at another hospital. The ITB decided on a completing surgery with ovariectomy on both sides, appendectomy as well as pelvic and paraaortic lymphadenectomy. ChatGPT failed to recognize the need for further surgical treatment. Another case worth discussing was patient 5, an 86-year-old woman with a number of concomitant diseases (including dementia, arterial hypertension and severe stress incontinence). Both the ITB and ChatGPT suggested surgical treatment. However, ChatGPT pointed out that surgery must be carefully considered in this patient due to her general state of health and advanced age. As an alternative to radical surgery, ChatGPT suggested less invasive surgical methods, such as laparoscopic removal of the tumor without further interventions on other organs. Another critical observation was ChatGPT’s tendency to omit salpingectomy and lymphadenectomy from surgical recommendations. However, hysterectomy and ovariectomy as well as peritoneal sampling and removal of all other affected tissue in the sense of a debulking operation were reliably mentioned by ChatGPT.

Chemotherapy and other systemic therapies. Chemotherapy was suggested by the ITB for four out of five ovarian cancer patients. The ITB did not recommend systemic therapy for patient 5 due to her known dementia. However, all of ChatGPT’s treatment plans for the patients with carcinoma of the ovaries covered chemotherapy. The AI was able to identify carboplatin and paclitaxel as suitable pharmaceutical agents for this treatment.

Other systemic therapies indicated by the ITB included VEGF antibody and PARP inhibitor therapy. ChatGPT recommended targeted therapy or immunotherapy for all patients with higher FIGO stages (IVA and B), in line with the German guidelines (5, 17). For one patient with ovarian cancer of FIGO stage IVA, the AI suggested maintenance therapy with bevacizumab after completion of chemotherapy with carboplatin and paclitaxel in three out of four treatment plans. PARP inhibitors were not explicitly mentioned once by ChatGPT. However, in two patients with FIGO stage IV ovarian cancer, the AI suggested testing the tumor tissue for specific mutations, in order to check the eligibility for targeted therapy. It should be mentioned that even the ITB did not consistently indicate PARP inhibitor therapy for patients with FIGO stage IV. This can be explained by the fact that genetic testing had not yet been carried out at the time of the treatment decision and the choice of a PARP inhibitor was directly dependent on this test result (5, 18, 19).

For a patient with low-grade serous ovarian carcinoma, both ITB and ChatGPT recommended endocrine therapy based on the hormone receptor status of the tumor. ChatGPT suggested postoperative hormone replacement therapy after bilateral adnexectomy for pre- and perimenopausal women only after careful consideration of the risks, reflecting current guideline ambiguities (5).

Genetic consulting. ChatGPT recommended genetic consulting to four out of five ovarian cancer patients, accurately identifying relevant genes like BReast CAncer 1/2 (BRCA 1/2). It appropriately did not suggest genetic testing for an 86-year-old patient with no family history, in line with current guidelines (20, 21). For two patients, genetic consulting was consistently advised in all treatment plans. For the remaining two, ChatGPT’s recommendations varied, potentially influenced by one patient’s explicit opposition to genetic testing, demonstrating its capacity to integrate patient preferences into its suggestions.

Further recommendations. Radiotherapy, typically not established in ovarian cancer treatment (5), was included in seven of ChatGPT’s 20 treatment plans. ChatGPT frequently suggested additional diagnostics, like supplemental imaging and tissue biopsies. It often specified biopsy methods, explicitly recommending surgical sampling and rightly advising against transdermal puncture for ovarian tumors due to metastasis risk (22). The AI also proposed cytology from peritoneal fluid in a case with ascites.

Cervical carcinoma. Answer scores for cervical carcinoma patients. In early-stage cervical carcinoma (CC) patients 6 and 7, ChatGPT performed relatively poorly with a total of five times (5 out of 8 AI-Rs) 0 points for incomplete recommendations (see Table II). For patients with advanced tumor stages (patients 8-11), ChatGPT’s recommendations were more frequently in line with those of the ITB (see Table II). A score of at least 1 was achieved in 12 out of 16 AI-Rs, and 0 points were awarded only four times (4 out of 16 AI-Rs). Across all queries for cervical carcinoma, ChatGPT achieved an average score of 0.7 points according to the answer scoring system.

Surgical treatment. Surgical resection of the tumor was suggested in two out of six cervical carcinoma patients by our ITB in accordance with the local guidelines. As one can see from Table II, there is a clear discrepancy with the answers from ChatGPT at this point. In patients 6 and 7, who suffered from early-stage cervical carcinoma (FIGO ≤ IA2), ChatGPT did not consistently suggest surgical treatment, in contrast to ITB.

Patient 6, for example, was a 34-year-old premenopausal woman with completed family planning. At the time of her initial presentation at our clinic she had already undergone laser conisation with cervical curettage and the histopathological examination revealed residual tumor tissue (R1). ChatGPT only recognized the need for further surgical intervention in two out of four AI-Rs.

Patient 7, a 27-year-old premenopausal woman, was presented to the ITB following a trachelectomy with bilateral sentinel node biopsy (histopathological tumor stage according to TNM 2017: pT1a2 pN0 L0 V0 Pn0 R0). According to the current German guidelines, a radical trachelectomy is an adequate treatment option in this tumor stage if the patient desires to have children and the lymph nodes are histologically negative (4). In this individual case, as patient 7 had not yet completed family planning, not suggesting any further surgical intervention should not be considered a mistake. However, it was not explicitly stated by ChatGPT that a secondary hysterectomy should be performed after a successful pregnancy, especially in cases of HPV persistence, PAP abnormalities, desire for safety, or limited evaluability of the cervix (4, 23).

Although the AI did not suggest adnexectomy for patients 6 and 7, ChatGPT proposed postoperative hormone replacement therapy for both women if necessary due to their premenopausal status.

Radiochemotherapy. In patient 8-11, our ITB recommended primary radiochemotherapy without prior tumor resection. ChatGPT regularly recognized the indication for radiochemotherapy and was able to identify cisplatin as a suitable radiosensitizer. In two patients, the AI additionally addressed the radiotherapeutic procedures of percutaneous radiotherapy and brachytherapy. Nevertheless, ChatGPT recommended additional radical hysterectomy in multiple AI-Rs (see Table II).

Lymphadenectomy. A lymphadenectomy in the form of a sentinel node biopsy was suggested by the ITB for one patient with early-stage cervical carcinoma (patient 6) due to the tumor type (adenosquamous cervical carcinoma), high grading (G3) of the tumor cells and the patient’s young age. According to the current German guidelines, there is no clear indication for sentinel node biopsy in the presence of an early tumor stage FIGO IA1 and lack of lymph vessel invasion (4). The recommendation for sentinel node biopsy can certainly be discussed in this case. However, not suggesting this procedure in the treatment plan is clearly not a mistake on ChatGPT’s part.

Furthermore, a lymph node examination was indicated by the ITB in patients 8, 10 and 11, each with advanced cervical carcinoma. It is important to note that patient 10 was suspected to have FIGO stage III preoperatively and was only diagnosed with FIGO stage IV during the operation. ChatGPT recognized the indication for surgical lymph node examination in the majority of queries (8 out of 12 AI-Rs).

ChatGPT did not recommend unnecessary lymphadenectomy in patients who had already undergone this procedure (patients 7 and 9), demonstrating its capacity to integrate prior medical interventions into its treatment plans.

Further recommendations. In the present study, ChatGPT demonstrated its versatility in treatment planning for advanced cervical carcinoma, repeatedly suggesting immunotherapy and delving into specific therapies like checkpoint inhibition with pembrolizumab. It also emphasized comprehensive patient care, advocating for preventive measures, such as HPV (human papilloma virus) vaccination, lifestyle changes (e.g., smoking and alcohol abstinence, healthy diet, exercise, stress management, sexual abstinence), and genetic consulting when relevant.

Endometrial carcinoma. Answer scores for endometrial carcinoma patients. For two patients, ChatGPT scored maximum points for every single AI-R. In the other three cases, the maximum answer score was achieved with at least one of the four treatment plans. The results are shown in Table III. Across all queries for endometrial carcinoma (EC), ChatGPT achieved an average score of 1.5 points according to the answer scoring system.

Surgical treatment and lymphadenectomy. In the presented data, surgical treatment recommendations from ChatGPT aligned closely with the ITB, suggesting surgical treatment for all EC patients, and emphasizing the preference for minimally invasive procedures if the oncological risk permitted. ChatGPT explicitly recommended hysterectomy in most cases (18 out of 20 AI-Rs). However, bilateral adnexectomy was suggested less consistently, in only nine AI-Rs, although all the endometrial carcinoma patients analyzed were postmenopausal and this intervention would have been indicated according to the German guidelines (16, 24). In three further AI-Rs, ChatGPT did recommend at least ovariectomy.

The need for lymphadenectomy was well identified by ChatGPT (see Table III), with the exception of patient 16. It is important to note that patient 16 was initially only known to have atypical endometrial hyperplasia and that the plans for surgical treatment - both from ChatGPT and the ITB - were based on this information. The diagnosis of endometrial cancer was not made until hysterectomy. This explains why ChatGPT did not indicate lymphadenectomy for this patient. In this case, the ITB did not clearly recommend a lymphadenectomy either. Instead, the patient was only offered a sentinel node biopsy. Eventually, the patient opted for lymph node surgery, which resulted in a negative nodal status [pN0 (0/3)]. ChatGPT’s deviation in recommending endometrial ablation as an alternative to hysterectomy for patient 16, contrary to guidelines, underscores the AI’s limitation in discerning suitable surgical interventions in complex cases (16).

Radiotherapy. According to the ITB, two of the patients were eligible for radiotherapy. ChatGPT aligned with this recommendation in three out of four therapy plans. However, it should be mentioned here that the AI classified radiotherapy in one of the treatment plans as “not absolutely necessary”, but merely as “to be considered”. For those patients for whom radiotherapy was not indicated by the ITB, ChatGPT also consistently did not suggest radiation.

Chemotherapy and other systemic therapies. Chemotherapy was recommended by the ITB for one out of five patients with endometrial cancer. For this patient, three out of four treatment plans generated by ChatGPT also included chemotherapy. Conversely, the AI suggested chemotherapy twice for a patient without a corresponding indication (see Table III). Other systemic therapies proposed by ChatGPT included immunotherapy and endocrine therapy. It should be mentioned in advance that these types of therapy were not recommended by the ITB even once in the patients examined. ChatGPT suggested immunotherapy for a total of two patients. In one of these patients, the tumor cells showed aberrant strong nuclear overexpression of p53 (p53 mutation signature). The AI presented the possibility of using p53 as a target for immunotherapy. In another patient with a first diagnosis of endometrial carcinoma FIGO stage IB with grading G3 and mismatch-repair deficiency, ChatGPT suggested considering immunotherapy (e.g., with immune checkpoint inhibitors), as mismatch repair (MMR)-deficient tumors may respond to this. According to the current German guideline, patients with relapsed or primarily advanced endometrial cancer with microsatellite-instable/mismatch-repair deficient tumor tissue can be treated with immunotherapy after prior treatment with platinum-based chemotherapy (16). Dostarlimab or pembrolizumab are particularly suitable substances for this therapy (16, 24). ChatGPT’s suggestions for immunotherapy are consequently not supported by the current guidelines. However, its therapy recommendations indicate that the AI is aware of certain correlations, for example between MMR deficiency and response to checkpoint inhibitor therapy (16, 24-26), and that ChatGPT incorporates this information into its treatment plans. Endocrine therapy was suggested more frequently by ChatGPT, in 11 out of 20 AI-Rs. In many cases, the type of hormonal therapy was not specified. In response to one query, the AI specifically recommended hormone replacement therapy in the postoperative setting after hysterectomy. A total of four times, endocrine therapy with progesterone was listed as a treatment option. ChatGPT identified megestrol acetate as an example of a suitable substance for such therapy. The AI emphasized that hormonal therapy is useful either in the adjuvant setting or if patients are not eligible for surgical treatment or radiotherapy.

Further recommendations. Another suggestion ChatGPT mentioned very frequently - more precisely in 15 out of 20 AI-Rs - was genetic consulting. In a patient with a p53 mutation of the tumor tissue and a family history of pancreatic cancer, the AI pointed out the possible presence of Li-Fraumeni syndrome. In another case where mutS homolog 6 (MSH6) loss of the tumor cells was detected, ChatGPT advised testing for Lynch syndrome. There was also one patient whose tumor showed an MMR protein deficiency of mutL homolog 1 (MLH1) and postmeiotic segregation increased 2 (PMS2). In this case, in addition to genetic consulting, the AI recommended an MLH1 promoter methylation analysis for further clarification. Other examinations ChatGPT occasionally requested included mainly imaging examinations for more precise staging (e.g. MRI examinations), which was the case for four patients in one AI-R each.

As can be seen from its answers, ChatGPT based its suggestions for diagnostics and therapy i.e., on immunohistochemical test results and, if these were not available, the AI actively requested them. In particular, subtyping of the tumor based on microsatellite instability, p53 and DNA polymerase epsilon (Polε) mutation status was required by ChatGPT multiple times. Using the molecular classification of endometrial carcinoma, the AI was also able to make statements about the prognosis of patients. For example, ChatGPT stated that Polε mutations can be associated with a more favorable prognostic profile.

Cross-entity observations. One observation for all carcinomas examined was that, depending on the patient’s history, the AI called for further diagnostics to be carried out pre-therapeutically, for example imaging with computer tomography or magnetic resonance imaging as well as clinical examination and determination of tumor markers. In the event of abnormalities in the staging, for example calcification in the liver, pulmonary lesions, splenic cysts, and nipple retraction, ChatGPT called for immediate diagnostic clarification or recommended a follow-up.

With regard to the indication of regular aftercare and follow-up examinations, ChatGPT was reliable in 93.75% of cases. For cervical carcinoma, the AI was able to correctly list the necessary examinations, for example gynecological examination, cytological smear test and, if necessary, HPV test, tumor marker controls, colposcopy, or imaging procedures. ChatGPT also provided details about the recommended check-ups for ovarian and endometrial cancer patients, including physical examinations, imaging, and tumor marker controls. In individual cases, ChatGPT gave additional information on the necessary follow-up intervals.

In ChatGPT’s responses, great importance was attached to supportive therapy, including analgesia, management of comorbidities as well as lifestyle adjustments, such as a balanced diet, regular physical activity, smoking cessation and abstaining from alcohol consumption. Depending on the patient’s complaints, also specific suggestions were made, for example pleural or ascites drainage. Interestingly, psychosocial support was suggested most frequently for ovarian cancer, in 65% of treatment plans (13/20 AI-Rs). In cervical cancer, this suggestion was made in 50% of cases (12/24 AI-Rs), and in endometrial cancer in only 20% (4/20 AI-Rs). One explanation for this could be that the patients with ovarian cancer in this study had on average a higher tumor stage compared to the patients of the CC and EC subgroup and therefore a significantly worse prognosis (27). It should be emphasized that these measures are part of the standard of care at our clinic, even without a prior tumor conference decision, and are therefore not regularly recorded in the ITB.

ChatGPT repeatedly emphasized the need for interdisciplinary care, tailored to the individual patient. In the case of an elderly patient with severe stress incontinence, for example, the AI suggested geriatric and palliative care as well as urological treatment in addition to gynecological-oncological therapy. ChatGPT’s consideration of clinical trial participation, especially for ovarian and cervical carcinoma cases, further underscores its potential in supporting nuanced cancer treatment strategies.